{"annotations": [{"id": 86721992, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86723280, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86723989, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19029-03cf4a9a918931f8.jsonl", "id_parag": "2310.19029-03cf4a9a918931f8_1", "final_parag": "\\textbf{Phase 1 (training)}: we recruited three undergraduate students majoring in linguistics. The students were trained in three steps in order to produce consistent annotations. We first assigned 50 words to each linguist and trained them to conduct the annotation jointly. Second, we assigned the same 150 words to each student separately, then asked them to compare and consolidate their annotations, which helps in calibrating their scoring. Third, we repeated the second phase, but using 300 words and again we asked them to compare their annotations.", "candidate_comments": ["annotator training and calibration, the annotation process and data quality and validation. \nTo ensure the accuracy and consistency of the annotations, it is essential carefully select and train the annotator linguists. In this subsection, we will describe the process we used to select and train the linguists for our corpus annotation task and how we evaluate these annotations. We started by recruiting a group of six linguists who were trained thoroughly on the annotation guidelines. After the training, we tested them carefully to evaluate their skills and select the best three linguists based on their performance. The selected linguists then carried out the sense-annotation process for our corpus. We conducted  the annotation process in three phases:"]}
{"annotations": [{"id": 86904180, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86981896, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86989533, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19029-03cf4a9a918931f8.jsonl", "id_parag": "2310.19029-03cf4a9a918931f8_2", "final_parag": "We evaluated the coverage of both lexicons based on the sense-annotated tokens. As Table \\ref{lexicons-evaluation} shows, Modern has higher coverage of lemmas (80\\%) compared to Ghani's coverage (78\\%), and has higher sense coverage (83\\%) compared to Ghani (78\\%). Moreover, glosses in Modern are more precise, less ambiguous and well-formulated as discussed in Section \\ref{sub:iaa-results}. The proper nouns are the main reason for the missing lemmas and senses, as the Modern and Ghani cover 4\\% and 14\\% of proper nouns in SALMA corpus, respectively. \nLemmas and senses that are not covered by any of the two lexicons were added manually by the linguists.", "candidate_comments": ["If a sense is missing, linguists can manually add a new sense in one of the lexicons; and when a lemma is missing, we add a new lemma and its senses manually. "]}
{"annotations": [{"id": 86738563, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86771982, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86826868, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19029-03cf4a9a918931f8.jsonl", "id_parag": "2310.19029-03cf4a9a918931f8_3", "final_parag": "where $fo_{ij}$ is the observed frequency of the categories ($i$ and $j$) per the annotators selection, $fe_{ij}$ is the expected frequency for both annotators' selected categories, $(y_{i}-y_{jx})$ denotes the distance between the categories, and $K$ is number of categories.", "candidate_comments": ["where $w_{ij}^x=\\frac{(y_{ix}-y_{jx})^2}{(k-1)^2}$, $y_{ix}$ and $y_{jx}$ denotes the score of gloss $x$ for a given token by annotator $i$ and $j$, respectively, and $k$ is number of categories. $fo_{ij}$ is the observed frequency of the categories ($i$ \\& $j$) per the annotators selection. $fe_{ij}$ is the expected frequency for both categories."]}
{"annotations": [{"id": 86722737, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86723424, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86723873, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19029-03cf4a9a918931f8.jsonl", "id_parag": "2310.19029-03cf4a9a918931f8_4", "final_parag": "Both LWK and QWK take the distance between categories into consideration, where the distance is defined as the number of categories separating the two annotators' selection. The difference is that LWK calculates the distance linearly while QWK calculates it quadratically. For measuring the ranking error deviation among annotators we used MAE and RMSE.", "candidate_comments": ["Kappa: data is nominal so no categories, the agreement is yes or no . threshold is >=60. \nWeighted Kappa: data is ordinal and measure the agreement based on the categories distance between annotators Linear Weighted: the distances between categories are equal  Quadratic Weithed: the distances scores between categories are not equal and getting higher penalty according to the importance level. \nthat is when the categories have difference importance. "]}
{"annotations": [{"id": 86961182, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86965601, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86970738, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19051-15a66630f02a71de.jsonl", "id_parag": "2310.19051-15a66630f02a71de_1", "final_parag": "\\subsubsection*{Acknowledgments} This work was supported in part by the National Natural Science Foundation of China under grant number 62167003, and in part by the Hainan Provincial Natural Science Foundation of China under grant number 720RC616.", "candidate_comments": [", in part by the Hainan Province Key R&D Program Project under grant number ZDYF2021GXJS010, and in part by the Major Science and Technology Project of Haikou City under grant number 2020006."]}
{"annotations": [{"id": 86962981, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86977232, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86977368, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_1", "final_parag": "Apple tasting is usually presented as an example of a more general partial feedback setting called \\textit{partial monitoring} games, where the player’s feedback is specified by a feedback matrix \\citep{cesa2006prediction, bartok2014partial}. Related to partial monitoring games is sequential prediction with \\textit{graph feedback}, for which apple tasting feedback is also special case \\citep{alon2015online}. In this model, a learner plays a repeated game against an adversary. In each round, the learner selects one of $K$ actions but observes the losses for a subset of the actions determined by a combinatorial structure called a \\textit{feedback graph}. \\cite{alon2015online} classify feedback graphs into three types and establish a trichotomy on the rates of the minimax regret based on the type of graph. In this paper, we extend the online learner presented in \\cite{alon2015online} to the setting of binary prediction with expert advice to establish the minimax regret of apple tasting in the agnostic setting. As mentioned in the introduction, \\cite{helmbold2000apple} were the first to study the online learnability under apple tasting feedback. In addition to establishing upper and lower bounds for apple tasting in terms of bounds on the false positive and false negative mistakes in the full-information setting, \\cite{helmbold2000apple} also give expected mistake bounds in terms of the size of $\\mathcal{H}$. More specifically, they show that for any hypothesis class $\\mathcal{H}$, there exists a randomized online learner whose expected number of mistakes, under apple tasting feedback, is at most $O\\biggl(\\min \\biggl\\{T, \\sqrt{\\frac{T \\ln(|\\mathcal{H}|)}{\\ln(1 + T/(\\ln(|\\mathcal{H}|)))}}, |\\mathcal{H}|\\biggl\\}\\biggl)$. They complement this result by giving a matching lower bound up to constant factor for classes that satisfy a property termed ``ample splitability\" (see Section 2.1 from \\cite{helmbold2000apple}).", "candidate_comments": ["In addition to establishing upper and lower bounds for apple tasting in terms of bounds on the false positive and false negative mistakes in the full-information setting, \\cite{helmbold2000apple} also give expected mistake bounds in terms of the size of $\\mathcal{H}$. More specifically, they show that for any hypothesis class $\\mathcal{H}$, there exists a randomized online learner whose expected number of mistakes, under apple tasting feedback, is at most $O\\biggl(\\min \\biggl\\{T, \\sqrt{\\frac{T \\ln(|\\mathcal{H}|)}{\\ln(1 + T/(\\ln(|\\mathcal{H}|)))}}, |\\mathcal{H}|\\biggl\\}\\biggl)$. They complement this result by giving a matching lower bound up to constant factor for classes that satisfy a property termed ``ample splitability\" (see Section 2.1 from \\cite{helmbold2000apple}). "]}
{"annotations": [{"id": 86960508, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86979199, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86979034, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_2", "final_parag": "\\subsection{Notation} Let $\\Xcal$ denote the instance space and $\\mathcal{H} \\subseteq \\{0, 1\\}^{\\mathcal{X}}$ denote a binary hypothesis class.", "candidate_comments": ["Given an instance $x \\in \\mathcal{X}$, and any collection of hypothesis $V \\subseteq \\{0, 1\\}^{\\mathcal{X}}$, we let $V(x) := \\{h(x): h \\in V\\}$ denote the projection of $V$ onto $x$."]}
{"annotations": [{"id": 86731004, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86726057, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86732169, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_3", "final_parag": "In the apple tasting feedback model, the adversary still picks a labeled instance $(x_t, y_t) \\in \\mathcal{X} \\times \\{0, 1\\}$ and reveals $x_t$ to the learner. However, the learner only gets to observe the true label $y_t$ if they predict $\\hat{y}_t = 1$. Analogous to the full-information setting, a hypothesis class $\\mathcal{H} \\subseteq \\{0, 1\\}^{\\Xcal}$ is online learnable under apple tasting feedback, if there exists an online learning algorithm whose expected regret, \\textit{under apple tasting feedback}, on any sequence of labeled instances is a sublinear function of $T$.", "candidate_comments": ["for any  chosen sequence of labeled examples $(x_t, y_t) \\in \\mathcal{X} \\times \\{0, 1\\}$, the algorithm, while only receiving feedback when predicting $1$, makes predictions $\\mathcal{A}(x_t) \\in \\{0, 1\\}$ at every iteration $t \\in [T]$ such that its \\emph{expected regret},"]}
{"annotations": [{"id": 86980402, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86984108, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86984842, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_4", "final_parag": "A hypothesis class $\\Hcal$ is online learnable under apple tasting feedback, if there exists an (potentially randomized) algorithm $\\mathcal{A}$ such that its \\emph{expected regret},", "candidate_comments": ["for any  chosen sequence of labeled examples $(x_t, y_t) \\in \\mathcal{X} \\times \\{0, 1\\}$, the algorithm, while only receiving feedback when predicting $1$, makes predictions $\\mathcal{A}(x_t) \\in \\{0, 1\\}$ at every iteration $t \\in [T]$ such that its \\emph{expected regret},"]}
{"annotations": [{"id": 86738693, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86901183, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86901328, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_5", "final_parag": "$$ \\inf_{\\mathcal{A}} \\texttt{M}_{\\mathcal{A}}(T, \\mathcal{H})=\\begin{cases} \\Theta(1), & \\text{if $\\texttt{w}(\\mathcal{H}) = 1$}\\\\ \\Theta(\\sqrt{T}) & \\text{if $2 \\leq \\texttt{w}(\\mathcal{H}) < \\infty$}\\\\ \\Theta(T), & \\text{otherwise} \\end{cases} $$", "candidate_comments": ["If $\\texttt{AL}_{\\texttt{w}(\\mathcal{H})}(\\mathcal{H}) > \\sqrt{(\\texttt{w}(\\mathcal{H}) - 1)T}$, then we have that $\\inf_{\\mathcal{A}}\\texttt{M}_{\\mathcal{A}}(T, \\mathcal{H}) \\leq 3\\texttt{AL}_{\\texttt{w}(\\mathcal{H})}$ while   "]}
{"annotations": [{"id": 86829935, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86861673, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86898739, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_6", "final_parag": "$$\\inf_{\\mathcal{A}}\\texttt{M}_{\\mathcal{A}}(T, \\mathcal{H}) \\leq \\inf_{w \\in \\mathbb{N}} \\left\\{\\texttt{AL}_w(\\mathcal{H}) + 2\\sqrt{(w-1)T} \\right\\}.", "candidate_comments": ["\\leq \\min\\Bigl\\{\\texttt{AL}_{\\texttt{w}(\\mathcal{H})}(\\mathcal{H}) + 2\\sqrt{(\\texttt{w}(\\mathcal{H})-1)T}, 2\\sqrt{\\texttt{L}(\\mathcal{H})T}\\Bigl\\}."]}
{"annotations": [{"id": 86961591, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972171, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86971725, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_7", "final_parag": "The upperbound in Theorem \\ref{thm:real} follows by picking $w = \\texttt{w}(\\mathcal{H})$ in Lemma \\ref{lem:up}. Note that if you pick $w = \\texttt{L}(\\mathcal{H})+1$, then $\\texttt{AL}_w(\\mathcal{H}) = \\texttt{L}(\\mathcal{H})$ and we get an upperbound of $3\\sqrt{\\texttt{L}(\\mathcal{H})T}$ on the expected mistakes.", "candidate_comments": ["\\leq \\min\\Bigl\\{\\texttt{AL}_{\\texttt{w}(\\mathcal{H})}(\\mathcal{H}) + 2\\sqrt{(\\texttt{w}(\\mathcal{H})-1)T}, 2\\sqrt{\\texttt{L}(\\mathcal{H})T}\\Bigl\\}."]}
{"annotations": [{"id": 86900883, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86901879, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86909644, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_8", "final_parag": "Lemma \\ref{lem:up} follows from composing the next two lemmas. Lemma \\ref{lem:soaconstrained} shows that if $\\texttt{AL}_w(\\mathcal{H}) < \\infty$, then there exists a deterministic online learner, under \\textit{full-information} feedback, that makes at most $w-1$ false negative mistakes and $\\texttt{AL}_w(\\mathcal{H})$ false positive mistakes. Lemma \\ref{lem:conver} is from \\cite{helmbold2000apple} and shows how to convert any online learner under full-information feedback into an online learner under apple tasting feedback.", "candidate_comments": ["Theorem \\ref{lem:up} implies that when $\\texttt{w}(\\mathcal{H}) = 1$, a constant upperbound on the expected regret is possible. In fact, when $\\texttt{AL}_1(\\mathcal{H}) < \\infty$, there exists a \\textit{deterministic} online learner which makes at most $\\texttt{AL}_1(\\mathcal{H}) $ mistakes in the realizable setting under apple tasting feedback (see Section \\ref{sec:updetreal} for more details). In Section \\ref{sec:lbrandomreal}, we give a matching lower bound showing that the upperbound given by Theorem \\ref{lem:up} is tight up to constant factors.  "]}
{"annotations": [{"id": 86899470, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86962132, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86963798, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_9", "final_parag": "\\begin{lemma} For any hypothesis class $\\mathcal{H}$ and $w \\in \\mathbbm{N}$ such that $\\texttt{AL}_w(\\mathcal{H}) < \\infty$, there exists a deterministic online learner which, under full-information feedback, makes at most $w-1$ \\emph{false negative} mistakes and at most $\\texttt{AL}_w(\\mathcal{H})$ \\emph{false positive} mistakes in the realizable setting. \\end{lemma}", "candidate_comments": ["Theorem \\ref{lem:up} implies that when $\\texttt{w}(\\mathcal{H}) = 1$, a constant upperbound on the expected regret is possible. In fact, when $\\texttt{AL}_1(\\mathcal{H}) < \\infty$, there exists a \\textit{deterministic} online learner which makes at most $\\texttt{AL}_1(\\mathcal{H}) $ mistakes in the realizable setting under apple tasting feedback (see Section \\ref{sec:updetreal} for more details). In Section \\ref{sec:lbrandomreal}, we give a matching lower bound showing that the upperbound given by Theorem \\ref{lem:up} is tight up to constant factors.  "]}
{"annotations": [{"id": 86961192, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86979305, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86979891, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_10", "final_parag": "The lowerbound in Theorem \\ref{thm:real} follows by picking $w = \\texttt{w}(\\mathcal{H}) - 1$ and and $w = \\texttt{L}(\\mathcal{H}) + 1$ respectively.", "candidate_comments": ["Moreover, compared to the upperbound given by Theorem \\ref{lem:up}, the lower bound given by Theorem \\ref{thm:lb} is tight up to an additive factor of $\\texttt{AL}_{\\texttt{w}(\\mathcal{H})}(\\mathcal{H})$.  "]}
{"annotations": [{"id": 86961583, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972522, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974059, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_11", "final_parag": "Let $\\mathcal{H} \\subseteq \\{0, 1\\}^{\\mathcal{X}}$, $w \\in \\mathbbm{N}$, and $T \\in \\mathbb{N}$ be the time horizon. Since learning under apple tasting feedback implies learner under full-information feedback, a lowerbound of $\\frac{\\min\\{T, \\texttt{L}(\\mathcal{H})\\}}{2}$ on the minimax expected mistakes follows trivially from the full-information feedback lowerbound. Accordingly, for the remainder of the proof we suppose $w \\leq \\text{L}(\\mathcal{H})$ and $T > w$, since if either condition is not met, the claimed lowerbound matches $\\frac{\\min\\{T, \\texttt{L}(\\mathcal{H})\\}}{2}$.", "candidate_comments": ["For all $w \\geq \\texttt{L}(\\mathcal{H}) + 1$, the stated lowerbound becomes $\\frac{1}{2}\\sqrt{\\texttt{L}(\\mathcal{H}) \\,\\min\\{T, \\texttt{L}(\\mathcal{H})\\}} \\leq \\frac{\\texttt{L}(\\mathcal{H})}{2}$, matching the full-information feedback lowerbound. Accordingly, suppose $w \\leq \\texttt{L}(\\mathcal{H})$."]}
{"annotations": [{"id": 86960147, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973407, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973554, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_12", "final_parag": "We first construct a path $\\sigma^{\\star}$ down $\\mathcal{T}$ recursively using $\\mathcal{A}$. Starting with $\\sigma^{\\star}_1$, let $A_1$ be the event that $\\mathcal{A}$, if presented with $\\frac{d}{w}$ copies of the root node $x^{\\star}_1$, predicts $1$ on at least one of the copies. Then, set $\\sigma^{\\star}_1 = 0$ if $\\mathbbm{P}(A_1) \\geq \\frac{1}{2}$ and set $\\sigma^{\\star}_1 = 1$ otherwise. For $j \\geq 2$, let $x^{\\star}_1, ..., x^{\\star}_j$ be the sequence of instances labeling the internal nodes along the prefix $(\\sigma^{\\star}_1, ..., \\sigma^{\\star}_{j-1})$ down $\\mathcal{T}$. Let $A_j$ be the event that $\\mathcal{A}$, if simulated with the sequence of $\\frac{d (j-1)}{w}$ labeled instances consisting of $\\frac{d}{w}$ copies of the labeled instance $(x^{\\star}_1, \\sigma^{\\star}_{1})$, followed by $\\frac{d}{w}$ copies of the labeled instance $(x^{\\star}_2, \\sigma^{\\star}_{2})$,..., followed by $\\frac{d}{w}$ copies of the labeled instance $(x^{\\star}_{j-1}, \\sigma^{\\star}_{j-1})$, predicts the label $1$ at least once when presented with $\\frac{d}{w}$ copies of the instance $x^{\\star}_j$. Set $\\sigma^{\\star}_j = 0$ if $\\mathbbm{P}(A_j) \\geq \\frac{1}{2}$ and set $\\sigma^{\\star}_j = 1$ otherwise. Continue this process until $\\sigma^{\\star}$ is a valid path that reaches the end of tree $\\mathcal{T}$.", "candidate_comments": ["Let $\\Sigma$ denote the set of all valid paths down $\\mathcal{T}$. Fix a path $\\sigma \\in \\Sigma$. Let $x_1, ..., x_{|\\sigma|}$ be the sequence of instances labeling the internal nodes along the path $\\sigma$ down $\\mathcal{T}$, where $x_1$ is the instance labeling the root node. We will define a sequence of events $A_{\\sigma} = (A^1_{\\sigma}, ..., A^{|\\sigma|}_{\\sigma})$ using $\\mathcal{A}$. Let $A^1_{\\sigma}$ be the event that $\\mathcal{A}$, if presented with $\\frac{d}{w}$ copies of $x_1$, predicts positively on at least one of the copies. For $2 \\leq j \\leq |\\sigma|$, let $A^j_{\\sigma}$ be the event that $\\mathcal{A}$, if simulated with the sequence of $\\frac{d (j-1)}{w}$ labeled instances consisting of $\\frac{d}{w}$ copies of the labeled instance $(x_1, \\sigma_{1})$, followed by  $\\frac{d}{w}$ copies of the labeled instance $(x_2, \\sigma_{2})$,..., followed by $\\frac{d}{w}$ copies of the labeled instance $(x_{j-1}, \\sigma_{j-1})$, predicts the label $1$ at least once when presented with $\\frac{d}{w}$ copies of the instance $x_j$. For every $\\sigma \\in \\Sigma$, let $A_{\\sigma}$ be the associated sequence of events as described above."]}
{"annotations": [{"id": 86961188, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972979, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86973442, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_13", "final_parag": "We now construct our hard labeled stream in blocks of size $\\frac{d}{w}$. Each block only contains a single labeled instance, repeated $\\frac{d}{w}$ times. For the first block $B_1$, repeat the labeled instance $(x_1^{\\star}, 0)$ if $\\sigma^{\\star}_1 = 0$ and otherwise repeat the labeled instance $(x_1^{\\star}, 1)$. Likewise, for block $B_j$ for $2 \\leq j \\leq |\\sigma^{\\star}|$, repeat for $\\frac{d}{w}$ times the labeled instance $(x_j^{\\star}, \\sigma^{\\star}_j)$. Now, consider the stream $S = (B_1, ..., B_{|\\sigma^{\\star}|})$ obtained by concatenating the blocks $B_1, ..., B_{|\\sigma^{\\star}|}$ in that order. If $\\frac{|\\sigma^{\\star}|d}{w} < T$, populate the rest of the stream $S$ with the labeled instance $(x^{\\star}_{|\\sigma^{\\star}|}, \\sigma^{\\star}_{|\\sigma^{\\star}|}).$", "candidate_comments": ["We will now use the events above to construct a hard stream for $\\mathcal{A}$ with the stated guarantee. Let $\\sigma^{\\star} \\in \\Sigma$ be the valid path such that for all $j \\in [|\\sigma^{\\star}|]$, if $\\sigma^{\\star}_j = 0$ then $\\mathbbm{P}(A^j_{\\sigma^{\\star}}) \\geq \\frac{1}{2}$ and if $\\sigma^{\\star}_j = 1$ then $\\mathbbm{P}(A^j_{\\sigma^{\\star}}) < \\frac{1}{2}$. Note that such a path is always guaranteed to exist. Let $x_1^{\\star}, ..., x_{|\\sigma^{\\star}|}^{\\star}$ be the sequence of instances labeling the internal nodes along the path $\\sigma^{\\star}$ down $\\mathcal{T}$. "]}
{"annotations": [{"id": 86725307, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86727072, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86765284, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_14", "final_parag": "We now lower bound the expected mistakes of $\\mathcal{A}$ on the entire stream $S$ by considering the number of ones in $\\sigma_{\\star}$ on a case by case basis. Note that since $\\sigma^{\\star}$ is a valid path down $\\mathcal{T}$, we have $w \\leq |\\sigma^{\\star}| \\leq d$. Consider the case where $\\sigma^{\\star}$ has $w$ ones. Then, $\\mathcal{A}$'s expected regret is at least its expected regret on those batches $B_j$ where $\\sigma^{\\star}_j = 1$. Thus, its expected regret is at least $w\\frac{d}{2w} = \\frac{d}{2}$. Consider the case where $\\sigma^{\\star}$ has $w-j$ ones for $w \\geq j \\geq 1$. Then, since $\\sigma^{\\star}$ is a valid path, it must be the case that there are $d - (w - j)$ zero's in $\\sigma^{\\star}$. Therefore, $\\mathcal{A}$'s expected regret is at least", "candidate_comments": ["We will now use the events above to construct a hard stream for $\\mathcal{A}$ with the stated guarantee. Let $\\sigma^{\\star} \\in \\Sigma$ be the valid path such that for all $j \\in [|\\sigma^{\\star}|]$, if $\\sigma^{\\star}_j = 0$ then $\\mathbbm{P}(A^j_{\\sigma^{\\star}}) \\geq \\frac{1}{2}$ and if $\\sigma^{\\star}_j = 1$ then $\\mathbbm{P}(A^j_{\\sigma^{\\star}}) < \\frac{1}{2}$. Note that such a path is always guaranteed to exist. Let $x_1^{\\star}, ..., x_{|\\sigma^{\\star}|}^{\\star}$ be the sequence of instances labeling the internal nodes along the path $\\sigma^{\\star}$ down $\\mathcal{T}$. "]}
{"annotations": [{"id": 86904730, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86985472, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86985645, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_15", "final_parag": "\\begin{thm}[EXP4.AT Regret Bound] If $\\eta = \\sqrt{\\frac{\\ln{N}}{2T}}$, then for any sequence of true labels $y_1, ..., y_T$, the predictions $\\hat{y}_1, ..., \\hat{y}_T$, output by EXP4.AT satisfy:", "candidate_comments": ["\\begin{lemma} For any $\\eta > 0$ and any sequence of true labels $y_1, ..., y_T$, the probabilities $p_1, ..., p_T$ output by EXP4.AT satisfy"]}
{"annotations": [{"id": 86962815, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973156, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86973844, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_16", "final_parag": "$$\\mathbbm{E}\\left[\\sum_{t=1}^T \\mathbbm{1}\\{y_t \\neq \\hat{y}_t\\} \\right] \\leq \\inf_{j \\in [N]} \\sum_{t=1}^T \\mathbbm{1}\\{y_t \\neq \\mathcal{E}^j_t\\} + 3\\sqrt{T\\ln{N}}.$$", "candidate_comments": ["$$\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\hat{\\ell}_t(y) - \\inf_{i \\in [N]}\\sum_{t=1}^T \\hat{\\ell}_t(\\mathcal{E}_t^i) \\leq \\frac{\\ln N}{\\eta} + \\eta \\sum_{t=1}^T \\hat{\\ell}_t(1) + \\eta\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\hat{\\ell}_t(y)^2.$$"]}
{"annotations": [{"id": 86865047, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86901885, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86903568, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_17", "final_parag": "$$\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\hat{\\ell}_t(y) - \\inf_{j \\in [N]}\\sum_{t=1}^T \\hat{\\ell}_t(\\mathcal{E}_t^{j} ) \\leq \\frac{\\ln N}{\\eta} + \\eta \\sum_{t=1}^T \\hat{\\ell}_t(1) + \\eta\\sum_{t=1}^T p_t^1(1 - p_t^1) \\hat{\\ell}_t(0)^2 + \\eta\\sum_{t=1}^T p_t^1 \\hat{\\ell}_t(1)^2.$$ \\end{lemma}", "candidate_comments": ["$$\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\hat{\\ell}_t(y) - \\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} \\hat{\\ell}_t(\\mathcal{E}_t^{j}) \\leq \\frac{\\ln N}{\\eta} + \\eta \\sum_{t=1}^T \\hat{\\ell}_t(1) + \\eta\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\hat{\\ell}_t(y)^2.$$ "]}
{"annotations": [{"id": 86904454, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972604, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86972967, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_18", "final_parag": "$$\\sum_{t=1}^T \\sum_{i=1}^N q_t^i \\ell^{\\prime}_t(\\mathcal{E}_t^i) \\leq \\sum_{t=1}^T \\ell^{\\prime}_t(\\mathcal{E}_t^j) + \\frac{\\ln N}{\\eta} + \\eta \\sum_{t=1}^{T} \\sum_{i=1}^N q_t^i (\\ell^{\\prime}_t(\\mathcal{E}_t^i))^2.$$", "candidate_comments": ["Next, observe that  $$ \\sum_{i=1}^N q_t(i) \\hat{z}_t(i) =  \\mathbb{E}_{i \\sim q_t}\\left[\\mathbb{E}_{y \\sim \\mathcal{E}_t^i}\\left[\\hat{\\ell}_t(y) \\right] \\right] = \\mathbb{E}_{\\hat{y}_t \\sim p_t}\\left[\\hat{\\ell}_t(\\hat{y}_t) \\right] $$"]}
{"annotations": [{"id": 86737279, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}, {"id": 86898875, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86899523, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_19", "final_parag": "$$\\sum_{t=1}^T \\sum_{i=1}^N q_t^i \\hat{\\ell}_t(\\mathcal{E}_t^i) \\leq \\sum_{t=1}^T \\hat{\\ell}_t(\\mathcal{E}_t^j) + \\frac{\\ln N}{\\eta} + \\eta \\sum_{t=1}^{T} \\sum_{i=1}^N q_t^i (\\ell^{\\prime}_t(\\mathcal{E}_t^i))^2.$$", "candidate_comments": ["Next, observe that  $$ \\sum_{i=1}^N q_t(i) \\hat{z}_t(i) =  \\mathbb{E}_{i \\sim q_t}\\left[\\mathbb{E}_{y \\sim \\mathcal{E}_t^i}\\left[\\hat{\\ell}_t(y) \\right] \\right] = \\mathbb{E}_{\\hat{y}_t \\sim p_t}\\left[\\hat{\\ell}_t(\\hat{y}_t) \\right] $$", "$$\\sum_{t=1}^T \\mathbb{E}_{\\hat{y}_t \\sim p_t}\\left[\\hat{\\ell}_t(\\hat{y}_t) \\right] -  \\sum_{t=1}^T \\mathbb{E}_{y \\sim \\mathcal{E}_t^{i^*}}\\left[\\hat{\\ell}_t(y)\\right] \\leq \\frac{\\ln N}{\\eta} + \\eta\\sum_{t=1}^T |\\text{supp}(p_t)|.$$ "]}
{"annotations": [{"id": 86899135, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86901956, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}, {"id": 86909247, "completed_by": {"id": 140156}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_20", "final_parag": "Next, observe that $$ \\sum_{i=1}^N q_t^i \\hat{\\ell}_t(\\mathcal{E}_t^i) = \\left(\\sum_{i=1}^N q_t^i \\mathcal{E}_t^i\\right) \\hat{\\ell}_t(1) + \\left(1 - \\sum_{i=1}^N q_t^i \\mathcal{E}_t^i\\right) \\hat{\\ell}_t(0) = \\frac{1}{1-\\eta}\\sum_{y \\in \\{0, 1\\}} p_t^y \\hat{\\ell}_t(y) - \\frac{\\eta}{1-\\eta} \\hat{\\ell}_t(1)$$", "candidate_comments": ["Next, observe that  $$ \\sum_{i=1}^N q_t(i) \\hat{z}_t(i) =  \\mathbb{E}_{i \\sim q_t}\\left[\\mathbb{E}_{y \\sim \\mathcal{E}_t^i}\\left[\\hat{\\ell}_t(y) \\right] \\right] = \\mathbb{E}_{\\hat{y}_t \\sim p_t}\\left[\\hat{\\ell}_t(\\hat{y}_t) \\right] $$", "$$\\sum_{t=1}^T \\mathbb{E}_{\\hat{y}_t \\sim p_t}\\left[\\hat{\\ell}_t(\\hat{y}_t) \\right] -  \\sum_{t=1}^T \\mathbb{E}_{y \\sim \\mathcal{E}_t^{i^*}}\\left[\\hat{\\ell}_t(y)\\right] \\leq \\frac{\\ln N}{\\eta} + \\eta\\sum_{t=1}^T |\\text{supp}(p_t)|.$$ "]}
{"annotations": [{"id": 86962042, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86977397, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86979004, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_21", "final_parag": "$$\\frac{1}{1-\\eta} \\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\hat{\\ell}_t(y) - \\frac{\\eta}{(1 - \\eta)} \\sum_{t=1}^T \\hat{\\ell}_t(1) \\leq \\sum_{t=1}^T \\hat{\\ell}_t(\\mathcal{E}_t^{j}) + \\frac{\\ln N}{\\eta} + \\frac{\\eta}{1-\\eta}\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\ell^{\\prime}_t(y)^2.$$", "candidate_comments": ["$$\\sum_{t=1}^T \\mathbb{E}_{\\hat{y}_t \\sim p_t}\\left[\\hat{\\ell}_t(\\hat{y}_t) \\right] -  \\sum_{t=1}^T \\mathbb{E}_{y \\sim \\mathcal{E}_t^{i^*}}\\left[\\hat{\\ell}_t(y)\\right] \\leq \\frac{\\ln N}{\\eta} + \\eta\\sum_{t=1}^T |\\text{supp}(p_t)|.$$ "]}
{"annotations": [{"id": 86722930, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}, {"id": 86723473, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}, {"id": 86724058, "completed_by": {"id": 140156}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_22", "final_parag": "$$\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\hat{\\ell}_t(y) - (1-\\eta) \\sum_{t=1}^T \\hat{\\ell}_t(\\mathcal{E}_t^{j}) \\leq \\frac{(1 - \\eta)\\ln N}{\\eta} + \\eta \\sum_{t=1}^T \\hat{\\ell}_t(1) + \\eta\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\ell^{\\prime}_t(y)^2$$ which further implies the guarantee:", "candidate_comments": ["$$\\sum_{t=1}^T \\mathbb{E}_{\\hat{y}_t \\sim p_t}\\left[\\hat{\\ell}_t(\\hat{y}_t) \\right] -  \\sum_{t=1}^T \\mathbb{E}_{y \\sim \\mathcal{E}_t^{i^*}}\\left[\\hat{\\ell}_t(y)\\right] \\leq \\frac{\\ln N}{\\eta} + \\eta\\sum_{t=1}^T |\\text{supp}(p_t)|.$$ ", "$$\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\ell^{\\prime}_t(y) - \\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} \\ell^{\\prime}_t(\\mathcal{E}_t^{j}) \\leq \\frac{\\ln N}{\\eta} + \\eta \\sum_{t=1}^T \\ell^{\\prime}_t(1) + \\eta\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\ell^{\\prime}_t(y)^2.$$ "]}
{"annotations": [{"id": 86962915, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86981567, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86981396, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_23", "final_parag": "$$ \\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\hat{\\ell}_t(y) - \\sum_{t=1}^T \\hat{\\ell}_t(\\mathcal{E}_t^{j}) \\leq \\frac{\\ln N}{\\eta} + \\eta \\sum_{t=1}^T \\hat{\\ell}_t(1) + \\eta\\sum_{t=1}^T p_t^1 (1 - p_t^1) \\hat{\\ell}_t(0)^2 + \\eta\\sum_{t=1}^T p_t^1 \\hat{\\ell}_t(1)^2$$", "candidate_comments": ["$$\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\ell^{\\prime}_t(y) - \\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} \\ell^{\\prime}_t(\\mathcal{E}_t^{j}) \\leq \\frac{\\ln N}{\\eta} + \\eta \\sum_{t=1}^T \\ell^{\\prime}_t(1) + \\eta\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\ell^{\\prime}_t(y)^2.$$ "]}
{"annotations": [{"id": 86734777, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86901068, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86901283, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19064-7fbe2739dfcbc2f2.jsonl", "id_parag": "2310.19064-7fbe2739dfcbc2f2_24", "final_parag": "$$\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\hat{\\ell}_t(y) - \\sum_{t=1}^T \\hat{\\ell}_t(\\mathcal{E}_t^{j}) \\leq \\frac{\\ln N}{\\eta} + \\eta \\sum_{t=1}^T \\hat{\\ell}_t(1) + \\eta\\sum_{t=1}^T p_t^0 p_t^1 \\hat{\\ell}_t(0)^2 + \\eta\\sum_{t=1}^T p_t^1 \\hat{\\ell}_t(1)^2.$$", "candidate_comments": ["$$\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\ell^{\\prime}_t(y) - \\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} \\ell^{\\prime}_t(\\mathcal{E}_t^{j}) \\leq \\frac{\\ln N}{\\eta} + \\eta \\sum_{t=1}^T \\ell^{\\prime}_t(1) + \\eta\\sum_{t=1}^T \\sum_{y \\in \\{0, 1\\}} p_t^y \\ell^{\\prime}_t(y)^2.$$ "]}
{"annotations": [{"id": 86732094, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86732236, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86766173, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_1", "final_parag": "In the Learning from Label Proportions (LLP) problem, the goal is to infer a classifier $f: \\mathcal{X} \\mapsto \\mathcal{Y}$ that maps items $\\X \\in \\mathcal{X}$ to labels $Y \\in \\mathcal{Y}$. LLP differs from a standard classification problem in that items do not come with individual labels. Instead, during the training phase, the items are provided in groups called `bags', and only the proportions of labels in each bag is available.", "candidate_comments": ["In the Learning from Label Proportions (LLP) problem, the goal is to infer a classifier $f: \\mathcal{X} \\mapsto \\mathcal{Y}$ that maps items $\\X \\in \\mathcal{X}$ to labels $Y \\in \\mathcal{Y}$.  LLP differs from a standard classification problem in that items do not come with individual labels. Instead, during the training phase, the items are provided in groups called `bags,' and only the proportion of positive labels in each bag is available."]}
{"annotations": [{"id": 86982830, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985509, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86985597, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_2", "final_parag": "Following the notation of \\cite{kddpaper2023}, we define an instance of the LLP problem over $N$ items taken from feature space $\\mathcal{X}$; each item is associated to one of $L \\leq N$ disjoint groups, called bags. A specific problem instance is given by a set of pairs $D = \\{(\\mathbf{x}_i, b_i), i = 1, \\dots, N\\}$ and a matrix $\\mathbf{P} \\in [0,1]^{L \\times C}$, with $\\mathbf{x}_i \\in \\mathcal{X}$ and $b_i \\in \\{1, \\dots, L\\}$. We assume that each item $\\mathbf{x}_i$ is associated to a label $y_i \\in \\{1, \\cdots, C\\}$, which is unknown. The labels are used to define $\\mathbf{P}$, where a row $\\mathbf{p_\\ell} \\in [0, 1]^{C}$ is the vector of proportions in bag $\\ell$, i.e., $p_{\\ell, c} = \\frac{|\\{i \\given b_i = \\ell, y_i = c\\}|}{|\\{i \\given b_i = \\ell\\}|}$. Then, the goal is to learn a classifier that maps the items from feature space $\\mathcal{X}$ to $\\{1, \\cdots, C\\}$. We assume the binary classification case, i.e., $C=2$, in all the experiments performed in this paper. However, all the methods presented here apply to any number of classes $C \\geq 2$.", "candidate_comments": ["Following the notation of \\cite{kddpaper2023}, we define an instance of the LLP problem over $N$ items taken from feature space $\\mathcal{X}$; each item is associated to one of $L \\leq N$ disjoint sets, called bags. A specific problem instance is given by a set of pairs $D = \\{(\\mathbf{x}_i, b_i), i = 1, \\dots, N\\}$ and a vector $\\mathbf{p} \\in [0,1]^L$, with $\\mathbf{x}_i \\in \\mathcal{X}$ and $b_i \\in \\{1, \\dots, L\\}$. We assume that each item $\\mathbf{x}_i$ is associated to a label $y_i \\in \\{0, 1\\}$, which is unknown. The labels are used to define $\\mathbf{p}$, where $p_\\ell$ is the proportion of 1's in bag $\\ell$, i.e., $p_\\ell = \\frac{|\\{i \\given b_i = \\ell, y_i = 1\\}|}{|\\{i \\given b_i = \\ell\\}|}$. Then, the goal is to learn a classifier that maps the items from feature space $\\mathcal{X}$ to $\\{0, 1\\}$."]}
{"annotations": [{"id": 86738436, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86901770, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86901868, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_3", "final_parag": "Given a dataset with known labels, in which items have been grouped into bags, it can be important to test which LLP variant the dataset represents. We assume that $\\X \\not \\indep Y$, which is necessary for there to be a non-trivial algorithm for classification. Hence, we need to perform five tests to identify the variant of a given dataset. Note that in the LLP setting, there are a variety of types of data. The feature vector $\\X$ is typically high-dimensional, $Y$ is a binary label (but can also be categorical, in which case $C > 2$), and $B$ is categorical.", "candidate_comments": ["Given a dataset with known labels, in which items have been grouped into bags, it can be important to test which LLP variant the dataset represents.  We assume that $\\X \\not \\indep Y$, which is necessary for there to be a non-trivial algorithm for classification. Hence, we need to perform five tests to identify the variant of a given dataset. Note that in the LLP setting, there are a variety of types of data. The feature vector $\\X$ is typically high-dimensional, $Y$ is a binary label, and $B$ is categorical."]}
{"annotations": [{"id": 86964479, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86972808, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973477, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_4", "final_parag": "Likewise, the problem of hyperparameter selection in LLP has only been considered in a small number of studies. The authors in \\cite{hernandez2019framework} proposed a $k$-fold based method that assign bags to folds in a way in which each fold have similar proportions in comparison to the entire dataset. Since this method uses a subset of entire bags to train the model and the rest of bags for validation, we call this method \\emph{full bag $k$-fold}.  Moreover, \\cite{kddpaper2023} proposed new methods to select hyperparameters in LLP: \\emph{split-bag bootstrap}, \\emph{split-bag shuffle}, and \\emph{split-bag $k$-fold}. We use all of these options in our proposed standardized evaluation method.", "candidate_comments": ["Likewise, the problem of hyperparameter selection in LLP has only been considered in a small number of studies.  The authors in \\cite{hernandez2019framework} proposed a $k$-fold based method that assign bags to folds in a way in which each fold have similar proportion of positive instances in comparison to the entire dataset. Since this method uses a subset of entire bags to train the model and the rest of bags for validation, we call this method \\emph{full bag $k$-fold}.   Moreover, \\cite{kddpaper2023} proposed new methods to select hyperparameters in LLP: \\emph{split-bag bootstrap}, \\emph{split-bag shuffle}, and \\emph{split-bag $k$-fold}.  We use all of these options in our proposed standardized evaluation method."]}
{"annotations": [{"id": 87251072, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87259386, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87267890, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_5", "final_parag": "In this section, we present methods to generate LLP datasets for different variants given a base (classification) dataset, a specified number of bags, bag proportions, and sizes of bags. We assume that each item in the base dataset is used once in the resulting LLP dataset, so a minimal condition for feasibility is that bag sizes sum to the dataset size, and expected bag proportions matches the global proportions.", "candidate_comments": ["In this section, we present methods to generate LLP datasets for different variants given a base (classification) dataset, a specified number of bags, bag proportions, and sizes of bags.  We assume that each item in the base dataset is used once in the resulting LLP dataset, so a minimal condition for feasibility is that bag sizes sum to the dataset size, and expected bag proportion matches the global proportion of positive labels.  %\\mec{GF: is this last sentence true?}"]}
{"annotations": [{"id": 86986273, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86986902, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86986918, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_6", "final_parag": "The input of the problem is as follows: a standard classification dataset $D_c = \\{(x_i, y_i), i = 1,\\dots,N\\}$, where the item-label pairs are i.i.d. observations of a random vector $(\\mathbf{X}, Y)$ with distribution $P_{\\mathbf{X}, Y} ( \\mathbf{x}, y)$; the number of classes $C$; the number of bags, $L$; the size of each bag, i.e., the vector $\\mathbf{s}$; the proportions of labels in each bag, i.e., the matrix $\\mathbf{P} \\in [0, 1]^{L \\times C}$, whose rows are stochastic; and a LLP variant, $\\nu$. The goal is to associate each $(x_i, y_i)$ to a bag $b_i$, in order to build a LLP instance, such that the number of bags is $L$, the bag proportions and bag sizes obtained are $\\mathbf{s}$ and $\\mathbf{P}$, respectively, and the dependence structure between items, labels, and bags is $\\nu$.", "candidate_comments": ["The input of the problem is as follows: a standard classification dataset $D_c = \\{(x_i, y_i), i = 1,\\dots,N\\}$, where the item-label pairs are i.i.d. observations of a random vector $(\\mathbf{X}, Y)$ with distribution $P_{\\mathbf{X}, Y} ( \\mathbf{x}, y)$; the number of bags, $L$; the size of each bag, i.e., the vector $\\mathbf{s}$; the proportion of positive labels in each bag, i.e.,  the vector $\\mathbf{p}$; and a LLP variant, $\\nu$. The goal is to associate each $(x_i, y_i)$ to a bag $b_i$, in order to build a LLP instance, such that the number of bags is $L$, the bag proportions and bag sizes obtained are $\\mathbf{s}$ and $\\mathbf{p}$, respectively, and the dependence structure between items, labels, and bags is $\\nu$."]}
{"annotations": [{"id": 86734101, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86773192, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86827432, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_7", "final_parag": "In other words, the goal of the problem is to obtain a data generation mechanism that allows us to sample from a, possibly unknown, probability distribution $Pr(B \\given \\X, Y)$ which respects the dependence structure imposed by a given LLP variant and the constraints imposed by the bag sizes and bag proportions. In the next sections, we present a way to obtain this sampling mechanism for each LLP variant presented in Table \\ref{tbl:taxonomy}.", "candidate_comments": ["In other words, the goal of the problem is to obtain a data generation mechanism that allows us to sample from a, possibly unknown, probability distribution $Pr(B \\given \\X, Y)$ which respects the dependence structure imposed by a given LLP variant and the constraints imposed by the bag sizes and bag proportions. In the next sections, we present a way to obtain this sample mechanism for each LLP variant presented in Table \\ref{tbl:taxonomy}. "]}
{"annotations": [{"id": 86982260, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985190, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86985010, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_8", "final_parag": "One should note that the desired sampling mechanism can not directly take into account the input matrix of proportions, $\\mathbf{P}$, since that would imply $Y \\notindep B$. As as consequence, any attempt to generate a Naive LLP dataset in which the components of $\\mathbf{P}$ differ significantly of the global proportions in the classification dataset is not feasible.", "candidate_comments": ["One should note that the desired sampling mechanism can not directly take into account the input vector of proportions, $\\mathbf{p}$, once that would imply $Y \\notindep B$. As as consequence, any attempt to generate a Naive LLP dataset in which the components of $\\mathbf{p}$ differ significantly of the global proportion of positive instances in the classification dataset is likely unfeasible."]}
{"annotations": [{"id": 86904649, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973208, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86974600, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_9", "final_parag": "Since $Pr(Y\\given B)$ is specified by the bag proportions ($\\mathbf{P}$), $Pr(B)$ can be computed from the bag sizes ($\\mathbf{s}$), and $Pr(Y)$ can be computed from the labels, the bag assignment rule $Pr(B \\given \\X, Y)$ can be obtained without needing to consider the features of items.", "candidate_comments": ["Since $Pr(Y\\given B)$ is specified by the bag proportions ($\\mathbf{p}$), $Pr(B)$ can be computed from the bag sizes ($\\mathbf{s}$), and $Pr(Y)$ can be computed from the labels, the bag assignment rule $Pr(B \\given \\X, Y)$ can be obtained without needing to consider the features of items."]}
{"annotations": [{"id": 86960559, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86976916, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86977267, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_10", "final_parag": "In order to overcome these issues, our first step is to cluster the items of the input dataset. The idea is to partition the items into $Q$ clusters ($1, \\dots, Q$) in a way that items of the same cluster share some similarity in the feature space. One could imagine simply using the clusters as bags. However, by doing so, there is no guarantee that the constraints related to bag sizes and proportions would be respected (or reasonably approximated). Instead, we use the clusters to simplify the feature space.", "candidate_comments": ["In order to overcome these issues, our first step is to cluster the items of the input dataset. The idea is to partition the items into $C$ clusters ($1, \\dots, C$) in a way that items of the same cluster share some similarity in the feature space. One could imagine simply using the clusters as bags. However, by doing so, there is no guarantee that the constraints related to bag sizes and proportions would be respected (or reasonably approximated). Instead, we use the clusters to simplify the feature space."]}
{"annotations": [{"id": 86865182, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86959575, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86961934, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_11", "final_parag": "To solve this, it helps to rewrite (\\ref{eq:intermediate-prob-derivation}) as a matrix equation: [EQUATION] where $\\mathbf{P}_{Y,B} \\in \\mathbb{R}^{C \\times L}$ represents $Pr(Y,B)$, $\\mathbf{P}_{Y,Z}\\in\\mathbb{R}^{C \\times Q}$ represents $Pr(Y, Z = z)$, and $\\mathbf{P}_{B|Z} \\in \\mathbb{R}^{Q \\times L}$ represents $Pr(B\\given Z = z)$. Note that $\\mathbf{P}_{Y,B}$ and $\\mathbf{P}_{Y,Z}$ can be easily obtained from the problem's input and from an initial cluster assignment computed over the data.", "candidate_comments": ["where $\\mathbf{P}_{Y,B} \\in \\mathbb{R}^{2 \\times L}$ represents $Pr(Y,B)$, $\\mathbf{P}_{Y,Z}\\in\\mathbb{R}^{2\\times C}$ represents $Pr(Y, Z = z)$, and $\\mathbf{P}_{B|Z} \\in \\mathbb{R}^{C\\times L}$ represents $Pr(B\\given Z = z)$.  Note that $\\mathbf{P}_{Y,B}$ and $\\mathbf{P}_{Y,Z}$ can be easily obtained from the problem's input and from an initial cluster assignment computed over the data."]}
{"annotations": [{"id": 86981077, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86985743, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86985764, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_12", "final_parag": "To generate an Intermediate problem instance we obtain (or approximate) $\\mathbf{P}_{B\\given Z}$ by solving $\\mathbf{P}_{B\\given Z} = \\argmin_{\\mathbf{A}\\in\\mathcal{A}}\\Vert\\mathbf{P}_{Y,B} - \\mathbf{P}_{Y,Z}\\mathbf{A}\\Vert_F$, where $\\Vert\\cdot\\Vert_F$ represents the Frobenius norm and the optimization domain $\\mathcal{A}$ is the set of all matrices in $\\mathbb{R}^{Q\\times L}$ whose rows are stochastic. A locally optimal solution can be found via projected gradient descent, as described in detail in the Supplemental Material (see Appendix \\ref{sec:appendix}). We then use the result of the optimization problem stated above, i.e., $Pr(B\\given Z = z)$, as a proxy for the bag assignment rule $Pr(B\\given\\X)$.", "candidate_comments": ["To generate an Intermediate problem instance we obtain (or approximate) $\\mathbf{P}_{B\\given Z}$ by solving $\\mathbf{P}_{B\\given Z} = \\argmin_{\\mathbf{A}\\in\\mathcal{A}}\\Vert\\mathbf{P}_{Y,B} - \\mathbf{P}_{Y,Z}\\mathbf{A}\\Vert_F$, where $\\Vert\\cdot\\Vert_F$ represents the Frobenius norm and the optimization domain $\\mathcal{A}$ is the set of all matrices in $\\mathbb{R}^{C\\times L}$ whose rows are stochastic.  A locally optimal solution can be found via projected gradient descent, as described in detail in the Supplemental Material (see Appendix \\ref{sec:appendix}).  We then use the result of the optimization problem stated above, i.e., $Pr(B\\given Z = z)$, as a proxy for the bag assignment rule $Pr(B\\given\\X)$."]}
{"annotations": [{"id": 86869076, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86901733, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86901935, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_13", "final_parag": "Again, let $Z$ be the random variable representing the cluster of a given item. Then, the goal is to find a three-dimensional array $\\mathbf{P}_{Z,Y,B} \\in \\mathbb{R}^{Q \\times C \\times L}$ that encodes $Pr(Z, Y, B)$. To find a suitable $\\mathbf{P}_{Z,Y,B}$, we rely on the fact that the marginals $Pr(Z)$, $Pr(Y)$, $Pr(B)$, $Pr(Z, Y)$, and $Pr(Y, B)$ can be computed from the problem's input and the initial cluster assignment. Thus, we start with a random initialization for $\\mathbf{P}_{Z,Y,B}$ which will almost surely meet the desired dependence structure (namely, that no factorization of the distribution is possible). We then use Iterative Proportional Fitting (IPF, \\cite{bishop2007discrete}) to obtain a joint distribution that respects the constraints imposed by the marginals. Since IPF consists only of rescaling fibers (rows, columns, etc) of the array, it does not introduce any new dependence structure.\\footnote{We used \\texttt{ipfn} (\\url{https://pypi.org/project/ipfn/}) to apply IPF to a three-dimensional table.}", "candidate_comments": ["Again, let $Z$ be the random variable representing the cluster of a given item. Then, the goal is to find a three-dimensional array $\\mathbf{P}_{Z,Y,B} \\in \\mathbb{R}^{C\\times 2 \\times L}$ that encodes $Pr(Z, Y, B)$. To find a suitable $\\mathbf{P}_{Z,Y,B}$, we rely on the fact that the marginals $Pr(Z)$, $Pr(Y)$, $Pr(B)$, $Pr(Z, Y)$, and $P(Y, B)$ can be computed from the problem's input and the initial cluster assignment.  Thus, we start with a random initialization for $\\mathbf{P}_{Z,Y,B}$ which will almost surely meet the desired dependence structure (namely, that no factorization of the distribution is possible).  We then use Iterative Proportional Fitting (IPF, \\cite{bishop2007discrete}) to obtain a joint distribution that respects the constraints imposed by the marginals.  Since IPF consists only of rescaling fibers (rows, columns, etc) of the array, it does not introduce any new dependence structure.\\footnote{We used \\texttt{ipfn} (\\url{https://pypi.org/project/ipfn/}) to apply IPF to a three-dimensional table.}"]}
{"annotations": [{"id": 86981659, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86983624, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86984280, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_14", "final_parag": "For each LLP variant, there are some limitations of the generation methods. For the Naive and Simple variants, the limitations are due to the dependence structure. In the Naive case, since the bags are independent of both items and labels, one can generate only datasets with proportions close to the global proportion. For the Simple variant, the global proportions still imposes constraints, but these are looser: only the expected proportions needs to match the global proportions.", "candidate_comments": ["For each LLP variant, there are some limitations of the generation methods. For the Naive and Simple variants, the limitations are due to the dependence structure. In the Naive case, since the bags are independent of both items and labels, one can generate only datasets with proportions close to the global proportion -- ie, each $\\mathbf{p}_i$ must be close to $P(Y=1)$ for the entire dataset. For the Simple variant, the global proportion still imposes constraints, but these are looser: only the expected proportion needs to match the global proportion $P(Y=1)$. "]}
{"annotations": [{"id": 86862765, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86958501, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86959105, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_15", "final_parag": "In the other hand, the main limitation of Intermediate and Hard generation methods is that they rely on a clustering assignment. For the Intermediate variant, each bag contains a mixture of the clusters. As a result the proportions and bag sizes that can be achieved by the generation process are limited to convex combinations of the clusters' proportions and sizes. This limitation is reflected in the fact that our solution for Intermediate is only locally optimal. For the Hard variant, this limitation is not severe, since we can also use the labels to generate the bags. However, the process continues to rely on a the choice of clustering assignment.", "candidate_comments": ["For each LLP variant, there are some limitations of the generation methods. For the Naive and Simple variants, the limitations are due to the dependence structure. In the Naive case, since the bags are independent of both items and labels, one can generate only datasets with proportions close to the global proportion -- ie, each $\\mathbf{p}_i$ must be close to $P(Y=1)$ for the entire dataset. For the Simple variant, the global proportion still imposes constraints, but these are looser: only the expected proportion needs to match the global proportion $P(Y=1)$. "]}
{"annotations": [{"id": 86962216, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86984898, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985100, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_16", "final_parag": "In the experimental setup, we are also careful with splitting the data between training and testing sets and evaluation metrics. Regarding the train/test split, after splitting the data, we have to assure that the proportions vector passed to the algorithm is the true proportions of the training data. For this reason, we recompute the proportions using the training data, reflecting more a real LLP case. In terms of evaluation metrics, when the ground truth labels are available, it is straightforward to use some supervised learning score, as accuracy or $F_1$-score. However, it is important to take into consideration the same aspects that are taken for the classification literature when it comes to choose the evaluation metric. \nAs a final result of all these considerations, we present a meta algorithm to evaluate LLP algorithms as Algorithm \\ref{alg:meta-alg-llp}. \\footnoteref{footnote}", "candidate_comments": ["\\gf{As a final result of all these considerations, we present a meta algorithm to evaluate LLP algorithms\\footnoteref{footnote}. Given an LLP dataset, a LLP algorithm, a hyperparameter selection strategy, and a space of hyperparameters, we start splitting the data between training and testing sets. Afterwards, we recompute the proportions vector using the training set, use a grid search to find the best combination of hyperparameters for the algorithm, refit the model using the whole training data and the best combination of hyperparameters, and measure the performance of the trained model in the testing set. Algorithm \\ref{alg:meta-alg-llp} (see Appendix \\ref{sec:appendix}) presents the steps of this meta algorithm. Although this meta algorithm looks straightforward, in the best of our knowledge there is no standardized evaluation methods for LLP.}"]}
{"annotations": [{"id": 86739411, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86901642, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86901717, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_17", "final_parag": "For each combination of dataset, algorithm, and hyperparameter selection strategy, we run 30 executions of the meta algorithm presented in \\S~\\ref{sec:standard-eval-methods} (see Algorithm \\ref{alg:meta-alg-llp}), using 75\\% of the data for training. We measure performance using the $F_1$-score of the learned model in the test set.", "candidate_comments": ["For each combination of dataset, algorithm, and hyperparameter selection strategy, we run 30 executions of the meta algorithm presented in \\S~\\ref{sec:standard-eval-methods}, using 75\\% of the data for training. We measure performance using the $F_1$-score of the learned model in the test set. "]}
{"annotations": [{"id": 86981701, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}, {"id": 86980798, "completed_by": {"id": 140156}, "result": ["Différent", "Réécriture"]}, {"id": 86981853, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}], "agreement": 77.77777777777777, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_18", "final_parag": "Our experimental setup spans a total of 1,440 experiments, i.e., 360 experiments for each hyperparameter selection strategy. We trained 1,440$\\times$30 = 43,200 models, ie, each algorithm was trained 8,640 times. To handle such large experimental setup, we used the resources from Boston University's Shared Computing Cluster (SCC) \\footnote{\\url{https://www.bu.edu/tech/support/research/computing-resources/scc/}}. \nTo compare different algorithms, we consider all executions for the 4 hyperparameter selection strategies (4$\\times$30 = 120 executions). Then, we use a $t$-test to determine statistically significant difference in means, starting from the case with highest mean. Figure \\ref{fig:best-algorithm-dataset-variant} shows the best algorithm defined by this procedure per base dataset and LLP variant.", "candidate_comments": ["For each combination of dataset, algorithm, and hyperparameter selection strategy, we run 30 executions of the meta algorithm presented in \\S~\\ref{sec:standard-eval-methods}, using 75\\% of the data for training. We measure performance using the $F_1$-score of the learned model in the test set. ", "In terms of comparing different hyperparameters selection strategies, we compared them by algorithm. Each algorithm gives a vote to the best(s) hyperparameters selection strategies, using the same procedure with the t-test as before. Then, we get the majority vote to define the best(s) hyperparameters selection strategies. Figure \\ref{fig:best-hyperparam-dataset-variant} shows the best hyperparameter selection methods defined by this procedure per base dataset and LLP variant."]}
{"annotations": [{"id": 86984271, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986101, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86986083, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19065-66c30bd111f2a070.jsonl", "id_parag": "2310.19065-66c30bd111f2a070_19", "final_parag": "Figure~\\ref{fig:best-algorithm-dataset-variant} shows that the superiority of various algorithms differs depending on the LLP variant and base dataset.  That is, there is no best algorithm across all variants and base datasets. For instance, we can see that the neural network DLLP is superior for the CIFAR-10 dataset, and its performance is dominant for Hard LLP. On the other hand, for the adult dataset, mean map methods dominate; but among those, AMM is superior in the Naive variant, while there is more variety among the best algorithms for the other variants. Additionally, Figure~\\ref{fig:effect-sizes} shows that the statistically significant differences in performance found between LLP algorithms are typically substantial.", "candidate_comments": ["Following \\cite{kddpaper2023}, split-bag strategies outperform full-bag strategy for the non Naive LLP variants. Moreover, as in the LLP algorithms comparison, we observe that the performance of the hyperparameters selection strategies changes depending on the base dataset and variant. Although split-bag has a better performance in non Naive variants, its performance is better in the Intermediate variants in comparison to the others."]}
{"annotations": [{"id": 86903975, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973250, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973983, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19077-66c30bd111f2a070.jsonl", "id_parag": "2310.19077-66c30bd111f2a070_1", "final_parag": "The network is shared by a set of packet types (flows) $\\PD:=\\{1,2,\\cdots,\\PKT\\}$. A packet type (flow) $j \\in \\PD$ is characterized by its source $s_j \\in \\V$, its destination $z_j \\in \\V$, an end-to-end deadline $d_j \\in \\mathbb N\\cup\\{0\\}$, and a weight $w_j \\in \\mathbb R^+$. Packets of different types arrive during a time horizon of length $\\TC$.", "candidate_comments": ["Packets of different types  arrive in this network during a time horizon of length $\\TC$. We use $\\PD:=\\{1,2,\\cdots,\\PKT\\}$ to denote the set of packet types. A packet of type $\\mathrm j \\in \\PD$ is characterized by its source $s_j \\in \\V$, its destination $z_j \\in \\V$, its relative deadline $d_j \\in \\mathbb N\\cup\\{0\\}$, and its weight $w_j \\in \\mathbb R^+$. "]}
{"annotations": [{"id": 86986351, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86987461, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}, {"id": 86987517, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19077-66c30bd111f2a070.jsonl", "id_parag": "2310.19077-66c30bd111f2a070_2", "final_parag": "\\subsection{\\texorpdfstring{\\Cref{randomized-scheduling-offline}: \\ALGOFFSPELLED}{}} We first introduce \\Cref{randomized-scheduling-offline}, a scheduling algorithm that probabilistically forwards packets at each time slot based on their \\textit{age} (time since their arrival), \\textit{type}, and their \\textit{current node} in the system. Specifically, \\Cref{randomized-scheduling-offline} forwards a type-$j$ packet, which is currently at node $v$, over an outgoing link $\\ell \\in \\AO v$ with a probability determined by forwarding variables $f_{j\\ell}^\\trl$, which depend on the packet type $j$ and the number of time slots $\\trl \\in \\{0,1,\\cdots, d_j\\}$ that the packet has been in the system so far (i.e., its age).", "candidate_comments": ["Specifically,  \\Cref{randomized-scheduling-offline} maintains a set of forwarding probabilities $f_{j\\ell}^\\trl$, at each node $v$, for $j \\in \\PD$, $\\ell \\in \\mathcal{O}(v)$, and $\\trl \\in \\{1,2,\\cdots, d_j\\}$. forwards packets probabilistically from their current node, choosing link $\\ell$, with probability determined by forwarding variables $f_{j\\ell}^\\trl$, which depend on the packet type $j$ and the number of time slots $\\trl \\in \\{1,2,\\cdots, d_j\\}$ the packet has been in the system following its arrival (referred to as ``age``).", "\\Cref{randomized-scheduling-offline} forwards packets probabilistically from their current node, choosing link $\\ell$, with probability determined by forwarding variables $f_{j\\ell}^\\trl$, which depend on the packet type $j$ and the number of time slots $\\trl \\in \\{1,2,\\cdots, d_j\\}$ the packet has been in the system following its arrival (referred to as ``age``). "]}
{"annotations": [{"id": 87255280, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87258817, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87317336, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19077-66c30bd111f2a070.jsonl", "id_parag": "2310.19077-66c30bd111f2a070_3", "final_parag": "To see that, recall that $x_{jk}$ is the probability of scheduling packet-type $j$ over route-schedule $k$, and forwarding variable $\\fl j \\ell \\tau$ can be interpreted as the fraction of type-$j$ packets scheduled over link $\\ell$ at age $\\tau$. Then, \\dref{ES-obj} and \\dref{ES:one-choice} are equivalent to \\dref{FS-obj} and \\dref{FS:one-choice}, respectively, since $\\sum_{k\\in K_j} x_{jk} \\equiv \\sum_{\\ell\\in \\AO{s_j}} \\fl j {\\ell} {\\FI}$, where, recall that $\\AO{s_{j}}$ denotes the outgoing links of $s_j$ in $\\EG$. Similarly, capacity constraint \\dref{ES-capacity} is equivalent to \\dref{FS:capacity} with $\\epsilon =0$. The remaining constraints \\dref{FS:time-conserv}, \\dref{FS:edge-cases}, \\dref{FS-postv} are the flow conservation laws. At the end, similarly to network flow problems, we can map the solution $\\{\\fl j \\ell \\trl\\}$ to a solution for $\\{x_{jk}\\}$ due to the flow decomposition theorem \\cite{goldberg1998beyond}. Note that in $\\FS$, we have reduced the capacity in constraint \\dref{FS:capacity} by a factor $(1+\\epsilon)$. It is easy to verify that solving \\FS, with $\\epsilon >0$, results in at most a factor $\\frac{1}{1+\\epsilon}$ reduction in the optimal objective value, and consequently, $W_{\\FS}\\geq \\frac{W_{\\EST}}{1+\\epsilon}$.", "candidate_comments": ["  The simplified \\LP $\\EST$ still has exponentially many variables (there are exponentially many route-schedules $k$). \\NEW{We resolve this, by reformulating \\EST through a change of variables, allowing us} to work with a small set of \\textit{forwarding variables}, and find the corresponding forwarding-based solution of $\\EST$. Analogous reformulations have been utilized to a great extent in network flow problems \\cite{magnanti1993network} (for more information, refer to \\Cref{flow-relation}). The conversion results in the following \\LP: \\NEW{ \\begin{subequations} \\end{subequations} } To see that, recall that $x_{jk}$ is the probability of scheduling packet-type $j$ over route-schedule $k$, and forwarding variable  $\\fl j \\ell \\tau$ defines type $j$'s probability of scheduling over link $\\ell$ at age $\\tau$ (as defined in \\Cref{randomized-scheduling-offline}). %Then we can modify $\\EST$ accordingly to obtain its flow-based description $\\FS$.    Then, for example, to obtain the new objective function, note that our definitions are consistent with $\\sum_{k\\in K_j} x_{jk} \\equiv \\sum_{\\ell\\in \\AO{s_j}} \\NEW{\\fl j {\\ell} {\\FI}}$, where, recall that $\\AO{s_{j}}$ denotes the adjacent outgoing links of $s_j$ in $\\EG$. At the end, similarly to network flow problems, we can map the solution $\\NEW{\\{\\fl j \\ell \\trl}\\}$ to a solution for \\NEW{$\\{x_{jk}\\}$} due to the flow decomposition theorem \\cite{goldberg1998beyond}.   \\NEW{ Note that \\dref{FSFC-obj}-\\dref{FSFC-final} is nearly identical to $\\FS$ (defined earlier in \\dref{FS-obj}-\\dref{FS-postv}), with slightly reduced capacities. However, it is easy to verify that solving \\FS, as opposed to \\dref{FSFC-obj}-\\dref{FSFC-final}, results in at most $\\frac{1}{1+\\epsilon}$ reduction in the optimal objective value of \\dref{FSFC-obj}-\\dref{FSFC-final}. Since \\dref{FSFC-obj}-\\dref{FSFC-final} is a reformulation of \\EST, it follows: $W_{\\FS}\\geq \\frac{W_{\\EST}}{1+\\epsilon}$. \n}"]}
{"annotations": [{"id": 86863755, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86900916, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86901079, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19077-66c30bd111f2a070.jsonl", "id_parag": "2310.19077-66c30bd111f2a070_4", "final_parag": "In our simulations, we assign a random reward uniformly chosen between $(0,1)$ for each source-destination pair and a deadline $10$ time slots to each packet. To further evaluate the impact of traffic intensity and the number of flows, we simulate two cases. In one case, we maintain all top $30$ source-destination pairs, whereas in the second experiment, we maintain a random subset of $15$ source-destination pairs.", "candidate_comments": ["In \\Cref{fig:comparison-a} we consider the randomized packet types considered earlier. In our first comparison we scale the traffic intensity and capacity of the networks. We observe that our algorithm maintains a growing advantage over GLS-FP for larger Link capacities and traffic capacities."]}
{"annotations": [{"id": 86736611, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86902160, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86903544, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19077-66c30bd111f2a070.jsonl", "id_parag": "2310.19077-66c30bd111f2a070_5", "final_parag": "Indeed, we have \\[ \\sum_{\\phi=1}^{\\PHASES} \\TC_\\phi \\epsilon/\\TCP \\overset{(a)}= \\sqrt{\\mu \\log(2\\PKT/\\epsilon) \\epsilon/\\TCP} \\sum_{\\phi=1}^{\\PHASES} \\sqrt{2}^{(\\phi-1)} \\overset{(b)}\\leq \\frac{\\sqrt{\\mu \\log(2\\PKT/\\epsilon) \\epsilon/\\TCP} }{\\sqrt \\epsilon (\\sqrt 2 -1 )} = \\frac{\\sqrt{ \\mu \\log(2 \\PKT/\\epsilon)/\\TC^\\prime}}{\\sqrt 2 -1}\\overset{(c)}\\leq \\epsilon. \\] $(a)$ is from definition of $\\epsilon_\\phi$ and $\\TC_\\phi$. In $(b)$, we used the fact that $\\sum_{\\phi=1}^{\\PHASES} \\sqrt{2}^{\\phi-1}=\\frac{\\sqrt {2}^{\\PHASES}-1}{\\sqrt 2 - 1} = (1/\\sqrt{\\epsilon}-1)/(\\sqrt 2 -1)\\leq \\frac {1}{\\sqrt \\epsilon (\\sqrt 2 -1)}$ since $2^{\\PHASES/2} \\sqrt{\\epsilon} = 1$. \nFinally, $(c)$ holds under the assumption that $\\TC> \\frac{1}{(\\sqrt 2 -1)^2} \\mu \\log(2 \\PKT/\\epsilon)/\\epsilon^2 + \\DMAX \\log (1/\\epsilon)$ which implies $\\TC^\\prime> \\frac{1}{(\\sqrt 2 -1)^2} \\mu \\log(2 \\PKT/\\epsilon)/\\epsilon^2$. \nTherefore, from \\dref{related-bound} and due \\dref{approximation-ratio-to-analyze}, we obtain a fraction $(1-6\\epsilon)(1-4\\epsilon)\\geq (1-10\\epsilon)$ of $\\RI$.", "candidate_comments": ["\\left(\\frac{2}{\\sqrt 2 -1}\\right)^2 \\mu \\log(\\log(1/\\epsilon) \\frac{2P}{\\epsilon})/\\epsilon^2$ we get the bound for $\\epsilon$"]}
{"annotations": [{"id": 86726704, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86729608, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86730712, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19077-66c30bd111f2a070.jsonl", "id_parag": "2310.19077-66c30bd111f2a070_6", "final_parag": "\\end{proof} \\section{General Distributions with stationary arrival rates} In this section we consider general distributions on the arrival processes $\\{\\ajt\\}$ with stationary arrival rates, i.e., $\\avgpacks_{j}^t \\equiv \\avgpacks_j$. First, we provide a preliminary definition which allows us to extend \\Cref{main-theorem-offline}. \n\\begin{definition} Consider a nonnegative integer-valued random variable $Z$ with distribution $\\mathcal D$, and suppose $Z$ is bounded by a constant $M$. Consider a decomposition of $Z$ as the sum of $N$ independent random variables $\\{Z_i: Z_{i} \\leq M_i\\}$: \\[ Z = \\sum_{i=1}^{N} Z_{i}. \n\\] We define the dependency-degree of the decomposition as $\\max _i M_i$. Then, define the dependency-degree of distribution $\\mathcal D$, as the minimum dependency-degree of all decompositions for $\\mathcal D$. \n\\end{definition}", "candidate_comments": ["\\begin{definition}  Consider a nonnegative integer-valued random variable $Y$ with distribution $\\mathcal D$, and suppose $Y$ is bounded by a constant $M$. Consider a decomposition of $Y$ as the sum of  $M$ (possibly dependent) Bernoulli random variables $\\{Y_{i}\\in \\{0,1\\}\\}$: \\[ Y = \\sum_{i=1}^{M} Y_{i}. \n\\] Such a decomposition is always possible\\footnote{For example one such decomposition is $Y=\\sum_{i=1}^{M} \\IND{Y\\geq i}$.}.  Define the dependency-degree of a decomposition, as the maximum number of variables $\\{Y_{i}\\}$ that depend on each other.  Then, define the dependency-degree of distribution $\\mathcal D$, as the minimum dependency-degree of all decompositions for $\\mathcal D$. \nWe write $\\textnormal{\\textrm{degree}}(Y)$. \n\\end{definition}"]}
{"annotations": [{"id": 86872685, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}, {"id": 86901245, "completed_by": {"id": 62306}, "result": ["Réécriture", "Différent"]}, {"id": 86901344, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}], "agreement": 77.77777777777777, "source_file": "2310.19077-66c30bd111f2a070.jsonl", "id_parag": "2310.19077-66c30bd111f2a070_7", "final_parag": "We now state \\Cref{main-theorem-offline-dependencies}, which generalizes \\Cref{main-theorem-offline}. \n\\begin{theorem} Consider stationary arrival processes $\\{a_{j}^t\\}$, with the property that each packet type's total traffic within a fixed time window, $\\sumtlast a_{j}^t$, has at most dependency degree $D$ (for all $j$ and $t$). Then, given $\\epsilon \\in (0,1/3)$, $\\ALGOFF$ provides $(1-3\\epsilon)$-approximation to $\\RI$ when $\\CMN \\geq 2 D \\left(\\frac{ 1+\\epsilon}{\\epsilon}\\right)^2 \\log (L/\\epsilon)$ and $\\TC\\geq \\frac{2 \\DMAX^2} {\\epsilon}$. \\end{theorem} The minimum link-capacity threshold for a given approximation ratio, now depends on the dependency degree of the distribution of the total arrivals for the packet type in a window of $(\\DMAX+1)$ consecutive time slots. In other words, a lower dependency-degree results in better performance guarantees, given $\\CMN$.", "candidate_comments": ["that can impact any link at any time. In other words, a lower dependency-degree results in better performance guarantees, given $\\CMN$. \n\\begin{theorem} \\NEW{Consider arrival processes $\\{a_{j}^t\\}$ such that each link's packet arrivals at any time slot can depend on a maximum of $(D-1)$ other packets},    given an $\\epsilon \\in (0,1/3)$, $\\ALGOFF_{(1+\\epsilon)^{-1}}$ provides $(1-3\\epsilon)$-approximation to $\\RI$ when $\\CMN \\geq 2 D \\left(\\frac{ 1+\\epsilon}{\\epsilon}\\right)^2 \\log (L/\\epsilon)$ and $\\TC\\geq \\frac{2 \\DMAX^2} {\\epsilon}$.   \\end{theorem}", "\\NEW{ \\begin{corollary}  If for each $t$, the arrivals of packet type $j$ are drawn from distribution $\\mathcal D_{j}$ with dependency-degree at most $D$, FBPF yields a $(1-3 \\epsilon)$-approximation when $\\CMN \\geq 2 D \\left(\\frac{ 1+\\epsilon}{\\epsilon}\\right)^2 \\log (L/\\epsilon)$ and $\\TC\\geq \\frac{2 \\DMAX^2} {\\epsilon}$.  \\end{corollary} }"]}
{"annotations": [{"id": 86983219, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86986615, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86986660, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19077-66c30bd111f2a070.jsonl", "id_parag": "2310.19077-66c30bd111f2a070_8", "final_parag": "\\begin{lemma} In the case that $\\sumtlast a_{j}^t$ has at most dependency degree $D$, for all $j$ and $t$, using the forwarding probabilities $\\{f_{j\\ell}^{\\tau \\star}\\}$ for scheduling packets (Lines \\ref{alg-packet-for}-\\ref{alg-packet-for-end} of \\Cref{randomized-scheduling-offline}), the probability of a packet being dropped is at most $\\epsilon$, if $\\CMN \\geq 2 D \\left(\\frac{ 1+\\epsilon}{\\epsilon}\\right)^2 \\log \\frac L \\epsilon$. \n\\end{lemma} \\begin{proof} The proof in \\Cref{small-drop-chance-lemma} leverages the independence of packet types. More specifically the proof for \\Cref{small-drop-chance-lemma} relied on \\Cref{concentration-lemma} to show that any link's capacity is unlikely to be exceeded. This was then used to obtain through a union bound the final result. We follow similar reasoning here, however when using \\Cref{concentration-lemma} we need to define a set of independent variables, which now cannot be the arrival or not of a packet on the link.", "candidate_comments": ["that can impact any link at any time. In other words, a lower dependency-degree results in better performance guarantees, given $\\CMN$. \n\\begin{theorem} \\NEW{Consider arrival processes $\\{a_{j}^t\\}$ such that each link's packet arrivals at any time slot can depend on a maximum of $(D-1)$ other packets},    given an $\\epsilon \\in (0,1/3)$, $\\ALGOFF_{(1+\\epsilon)^{-1}}$ provides $(1-3\\epsilon)$-approximation to $\\RI$ when $\\CMN \\geq 2 D \\left(\\frac{ 1+\\epsilon}{\\epsilon}\\right)^2 \\log (L/\\epsilon)$ and $\\TC\\geq \\frac{2 \\DMAX^2} {\\epsilon}$.   \\end{theorem}"]}
{"annotations": [{"id": 86904259, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86952483, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86959197, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19077-79afd71960e8496f.jsonl", "id_parag": "2310.19077-79afd71960e8496f_1", "final_parag": "In order to evaluate the performance of our algorithms in practical settings, in addition to extensive synthetic simulations, we provide simulations using real traffic traces, and over real networks. \nThe results indicate that, despite the presence of highly non-stationary traffic in the network traces, our algorithms demonstrate significant performance improvement over the worst-case-based algorithms.", "candidate_comments": ["In order to assess the efficacy of our algorithms in realistic scenarios, we complement our comprehensive synthetic simulations with simulations utilizing actual traffic traces."]}
{"annotations": [{"id": 86962697, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86976610, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86979014, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19080-0acee387fce24dbd.jsonl", "id_parag": "2310.19080-0acee387fce24dbd_1", "final_parag": "\\centering \\caption{Analysis on rewards of boxes produced by different detectors. We report mean and std of box reward on the \\lyft dataset. }", "candidate_comments": ["\\centering \\caption{Analysis on rewards of boxes produced by different detectors. We report mean and std of box reward on the \\lyft dataset. }"]}
{"annotations": [{"id": 86872809, "completed_by": {"id": 70661}, "result": ["Réécriture", "Différent"]}, {"id": 86907852, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86903978, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}], "agreement": 55.55555555555556, "source_file": "2310.19081-66c30bd111f2a070.jsonl", "id_parag": "2310.19081-66c30bd111f2a070_1", "final_parag": "\\item \\textbf{Linear-frequency power spectrogram:} \\begin{comment} Represents the time on the x-axis, the frequency in Hz on a linear scale on the y-axis, and the power in dB \\cite{bib21p0}. \nIn audio forensics, a linear-frequency power spectrogram is used for various purposes, including: Identification of audio events by analyzing the spectrogram, audio forensic experts can identify specific events or occurrences within an audio recording, Voice analysis via Spectrograms help in analyzing the characteristics of human voices, such as pitch, formants, and harmonics. This can be useful for speaker identification, voice comparison, and determining whether a particular voice in a recording matches a known individual, Audio enhancement: Spectrograms can reveal certain artefacts, noise, or disturbances in an audio recording that are not easily detectable in the time-domain waveform. By identifying specific frequency ranges affected by noise or interference, audio forensic experts can apply appropriate filtering techniques to enhance the desired audio content, in the Audio authentication field it can be used to detect signs of audio tampering or manipulation and in the Forensic transcription, it can assist in the transcription process, especially when dealing with challenging or degraded audio recordings. Visualizing the frequency content helps in distinguishing and understanding speech components, even in difficult listening conditions and by examining the spectrogram, experts can identify irregularities, such as added or removed sections, pitch modifications, or other alterations, which may indicate potential tampering. \nOverall, the linear-frequency power spectrogram is a valuable tool in audio forensics, enabling experts to analyze and interpret the spectral characteristics of an audio recording, identify relevant events, enhance audio quality, detect tampering, and aid in transcription and voice analysis. \\end{comment} The linear-frequency power spectrogram is an important tool in the field of audio forensics. It represents time on the X-axis, frequency in Hz on a linear scale on the Y-axis, and power in dB \\cite{bib21p0}. It is used to identify specific events in an audio recording by allowing experts to analyze the spectral characteristics. It also enables voice analysis by studying features such as pitch, formants, and harmonics, which can aid in speaker identification and voice comparison. Furthermore, the spectrogram can reveal hidden artefacts, noise, or disturbances in an audio recording, which can then be mitigated by applying appropriate filtering techniques, thereby enhancing the desired audio content. In audio authentication, the spectrogram can be instrumental in detecting signs of audio tampering or manipulation.", "candidate_comments": ["To conclude, the linear-frequency power spectrogram enables experts to interpret spectral characteristics of an audio recording, identify relevant events, enhance audio quality, detect tampering, and provide assistance in transcription and voice analysis.", "A log-frequency power spectrogram is a valuable tool in audio forensics for speaker identification, audio classification, segmentation, transcription, and authentication. "]}
{"annotations": [{"id": 86904233, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86973366, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}, {"id": 86974249, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19081-66c30bd111f2a070.jsonl", "id_parag": "2310.19081-66c30bd111f2a070_2", "final_parag": "\\item \\textbf{Log-frequency power spectrogram:} Such features can be obtained from a spectrogram by converting the linear frequency axis (measured in Hertz) into a logarithmic axis (measured in pitches).", "candidate_comments": ["To conclude, the linear-frequency power spectrogram enables experts to interpret spectral characteristics of an audio recording, identify relevant events, enhance audio quality, detect tampering, and provide assistance in transcription and voice analysis.", "A log-frequency power spectrogram is a valuable tool in audio forensics for speaker identification, audio classification, segmentation, transcription, and authentication. "]}
{"annotations": [{"id": 86725034, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}, {"id": 86731507, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86827742, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}], "agreement": 55.55555555555556, "source_file": "2310.19081-66c30bd111f2a070.jsonl", "id_parag": "2310.19081-66c30bd111f2a070_3", "final_parag": "Its logarithmic representation of frequency content enables experts to extract unique voice characteristics, classify sounds, segment audio recordings, enhance transcription accuracy, and detect potential tampering or manipulation. \nThe resulting representation is also called log-frequency spectrogram.", "candidate_comments": ["To conclude, the linear-frequency power spectrogram enables experts to interpret spectral characteristics of an audio recording, identify relevant events, enhance audio quality, detect tampering, and provide assistance in transcription and voice analysis.", "A log-frequency power spectrogram is a valuable tool in audio forensics for speaker identification, audio classification, segmentation, transcription, and authentication. "]}
{"annotations": [{"id": 86981051, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985818, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86985907, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19081-66c30bd111f2a070.jsonl", "id_parag": "2310.19081-66c30bd111f2a070_4", "final_parag": "Chroma STFT features are useful for analyzing the harmonic content of an audio signal and can be used in a variety of applications such as music information retrieval, audio classification, and speech recognition. They provide a way to represent the pitch content of an audio signal in a compact and efficient way and can be used to compare and classify different audio signals based on their harmonic content. Chroma STFT is a useful tool in audio signal processing for analyzing the chromatic content of an audio signal and can be used in a wide range of applications. This implementation is derived from chromagram E \\cite{bib21p1}", "candidate_comments": ["Chroma STFT is a type of feature extraction method used in audio signal processing that represents the chroma content of an audio signal in the time-frequency domain. The chroma content of an audio signal is the set of 12 pitch classes that are commonly used in Western music (i.e., the 12 notes of the chromatic scale). \nTo compute a Chroma STFT, the audio signal is first windowed into short segments, and the short-time Fourier transform (STFT) is applied to each segment to obtain a time-frequency representation of the signal. Then, the magnitude of each STFT coefficient is weighted by a set of filter bank coefficients that correspond to the 12 pitch classes. The weighted magnitudes are summed over each pitch class to obtain a set of 12 Chroma features for each time frame."]}
{"annotations": [{"id": 86962493, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86971544, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86971652, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19081-66c30bd111f2a070.jsonl", "id_parag": "2310.19081-66c30bd111f2a070_5", "final_parag": "The choice of DCT basis functions depends on the specific application and the trade-offs between computational efficiency, frequency resolution, and energy compaction. In many cases, the standard DCT-II is a good choice for audio signal processing applications, but other DCT bases may be more appropriate for certain types of signals or processing tasks.", "candidate_comments": ["Here, we will discuss comparing different DCT bases. \nThe most commonly used DCT basis functions are the DCT-II, DCT-III, and DCT-IV.  The DCT-II is also known as the \"standard\" DCT and is commonly used for audio compression, while the DCT-III and DCT-IV are less widely used but have certain advantages for some applications. \nWhen comparing different DCT bases, it is important to consider factors such as computational efficiency, frequency resolution, and energy compaction. In general, the DCT-II is computationally efficient and provides good energy compaction for audio signals, making it a popular choice for many applications. However, the DCT-III and DCT-IV have some advantages over the DCT-II in certain contexts. For example, the DCT-III has better symmetry properties, which can be useful for certain types of signal processing, while the DCT-IV has better frequency resolution for some types of signals."]}
{"annotations": [{"id": 86961028, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972097, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86972297, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19081-66c30bd111f2a070.jsonl", "id_parag": "2310.19081-66c30bd111f2a070_6", "final_parag": "Suppose we wanted to test the quality of the neural networks available in the automatic speech recognition application for a specific language, using files not belonging to datasets. It is possible to do this by the pipeline creation section. \nThe user creates a pipeline and add as many steps as necessary to compare the neural networks available for that language and save the pipeline. \nThen the users can run the various comparison pipelines (previously created) to test the behaviour of the various networks for tasks in examples that are not included in the training datasets. \nIn this use case, it is not possible to perform a validation according to the metrics related to the task being analysed, because the relevant ground-thoughts are missing. For this reason, it was decided to predefine a perceptual quality index ranging from 1 to 10 for the tasks on the platform.", "candidate_comments": ["Using the Deep Audio Analyzer, it was possible to create different types of pipelines in a few steps to test the different state-of-the-art neural networks for each type of task, using files not belonging to datasets. In particular, "]}
{"annotations": [{"id": 86986662, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86987223, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86987324, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19081-66c30bd111f2a070.jsonl", "id_parag": "2310.19081-66c30bd111f2a070_7", "final_parag": "\\subsection{Results} \\subsubsection{Model Evaluation on different Datasets} Tables \\ref{tab:EvaluationASR}, \\ref{tab:EvaluationSpeechSeparation} show the Evaluation module applied Automatic Speech recognition task and Speech Separation task with pre-trained models on some datasets. However, evaluations conducted on different datasets show that even though a network may show good performance on the training dataset, it may not perform well on other data from different contexts. \nWith Deep Audio Analyzer is possible to upload customized trained models in order to achieve better performance on private datasets.", "candidate_comments": ["Using the Deep Audio Analyzer, it was possible to create different types of pipelines in a few steps to test the different state-of-the-art neural networks for each type of task, using files not belonging to datasets. In particular, "]}
{"annotations": [{"id": 87251223, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87259490, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87267771, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19081-66c30bd111f2a070.jsonl", "id_parag": "2310.19081-66c30bd111f2a070_8", "final_parag": "\\section{Conclusion } In this paper, we described Deep Audio Analyzer, an audio analysis platform that aims to cover the entire audio analysis process. Deep Audio Analyzer is a framework that allows the comparison of state-of-the-art models for speech analysis with no lines of code. It enables researchers to reduce the benchmarking time of different models.", "candidate_comments": ["\\subsection{Conclusion} Deep Audio Analyzer is a framework that allows to compare state-of-the-art models for speech analysis with no lines of code. This framework enables researchers to reduce the time enabling rapid benchmarking of different models used. In addition, through the pipeline creation feature, it is possible to create new audio analysis processes and evaluate different models for a given task and export it. %en data that are not part of any dataset and that are not evaluable by any ground truth. Once the pipeline has been created, the user can export it and share it with other users in order to replicate the same experiments by performing a simple import. %The pipelines created or imported by the user are available in the appropriate module, where it is possible to run them on single files or entire folders."]}
{"annotations": [{"id": 86904520, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86964379, "completed_by": {"id": 140156}, "result": ["Réécriture", "Différent"]}, {"id": 86966006, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}], "agreement": 77.77777777777777, "source_file": "2310.19083-00514a3d604be5ad.jsonl", "id_parag": "2310.19083-00514a3d604be5ad_1", "final_parag": "The computations times are $3.8\\si{\\second}$, $3.4\\si{\\second}$, and $1.9\\si{\\second}$ for the three cases. \nAs expected, the projections show that $\\outerBRSAEsuper{-\\tau}{1} \\supset \\outerBRSAEsuper{-\\tau}{2} \\supset \\outerBRSAEsuper{-\\tau}{3}$ since the input capacity increases via $\\zeta^{(1)} \\leq \\zeta^{(2)} \\leq \\zeta^{(3)}$ and the size of the disturbance set $\\distset{}$ decreases as $\\varphi^{(1)} \\geq \\varphi^{(2)} \\geq \\varphi^{(3)}$. \nIn the leftmost projection, we see that $\\outerBRSAEsuper{-\\tau}{1}$ extends furthest in $\\pm x_1$ and $\\pm x_3$ because the disturbance $w_1$ is larger than in the other cases and forces more states to enter the target set. \nAs indicated by the middle and rightmost plots, an increase of the input capacity of $u_1$ for $\\outerBRSAEsuper{-\\tau}{3}$ allows more states to avoid the target set $\\targetset{}$ in comparison to $\\outerBRSAEsuper{-\\tau}{2}$, which is affected by the same disturbance set. \nAlso note that the leftmost projections of $\\outerBRSAEsuper{-\\tau}{2}$ and $\\outerBRSAEsuper{-\\tau}{3}$ are identical because the input $u_1$ neither directly nor indirectly influences these dimensions. \nMoreover, the middle plot shows that no state with positive vertical velocity $x_4$ that is unable to avoid the target set.", "candidate_comments": ["In contrast, the projection onto the axes $x_2$-$x_4$ is equal for $\\outerBRSAEsuper{-\\tau}{1}$ and $\\outerBRSAEsuper{-\\tau}{1}$, as these dimensoins are only affected by the disturbance $w_2$ that is equal in both cases. \nIn the third case, the doubled input capacity allows more states to avoid the target set $\\targetset{}$, which decreases the extension of $\\outerBRSAEsuper{-\\tau}{3}$ along all dimensions that are affected by the control input $u_1$.", "Although the time horizon was set to $\\tau = [0,0.5]$, the computation aborts prematurely since the control input has enough time to steer all states away from the target set after a certain time despite worst-case disturbances. \nFor $\\outerBRSAEsuper{-\\tau}{1}$ and $\\outerBRSAEsuper{-\\tau}{2}$, the latest point in time is $t = 0.101$, whereas this time reduces to $t = 0.052$ for $\\outerBRSAEsuper{-\\tau}{3}$ due to changing the input capacity of $u_1$ via doubling the scaling factor $\\zeta$."]}
{"annotations": [{"id": 86961821, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86971383, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86971630, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-00514a3d604be5ad.jsonl", "id_parag": "2310.19083-00514a3d604be5ad_2", "final_parag": "First of all, we notice that the backward reachable sets are symmetric with respect to the origin along some dimensions, which is caused by the symmetry of the input set and the disturbance set---except for $u_{(1)} \\in [-9.81,2.38]$, which becomes apparent in the projection onto the $x_3$-$x_6$ axes. \nThe sets $\\innerBRSEAsuper{-\\tau}{1}$ and $\\innerBRSEAsuper{-\\tau}{2}$ are computed without any disturbance. \nHence, the backward reachable set $\\innerBRSEAsuper{-\\tau}{2}$ encloses $\\innerBRSEAsuper{-\\tau}{1}$ in the first two projections from the left since the contributing part of the input set $\\inputset{}$ is larger: $i \\in \\{1,2\\}: \\zeta^{(1)}_{(i)} > \\zeta^{(2)}_{(i)}$. \nIn constrast, the opposite containment holds true for the rightmost projection as $\\zeta^{(1)}_{(3)} > \\zeta^{(2)}_{(3)}$. \nThis follows our expectation because more input capacity can steer additional states into the target set, thereby enlarging the maximal backward reachable set.", "candidate_comments": ["Please also note that the projection onto the $x_3$-$x_6$ axis looks the same for $\\innerBRSEAsuper{-\\tau}{1}$ and $\\innerBRSEAsuper{-\\tau}{2}$ since these dimensions are only affected by the first input $u_{(1)}$, which is the same in both cases."]}
{"annotations": [{"id": 86983189, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86985566, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86985598, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-00514a3d604be5ad.jsonl", "id_parag": "2310.19083-00514a3d604be5ad_3", "final_parag": "The computation of $\\innerBRSEAsuper{-\\tau}{3}$ takes a non-zero disturbance into account, but also increases the input capacity compared to $\\innerBRSEAsuper{-\\tau}{2}$: The first three projections show a much smaller backward reachable set as the additional input capacity is outweighed by the disturbance. \nIn contrast, the relatively small disturbance $\\varphi^{(3)}_{(3)}$ does not affect the reachable set size in the rightmost projection too much but results in a slightly rotated set in comparison with $\\innerBRSEAsuper{-\\tau}{1}$ in combination with the increased input capacity. \nIn summary, this example demostrates well how different input capacities and disturbances affect the size of the maximal reachable set.", "candidate_comments": ["Please also note that the projection onto the $x_3$-$x_6$ axis looks the same for $\\innerBRSEAsuper{-\\tau}{1}$ and $\\innerBRSEAsuper{-\\tau}{2}$ since these dimensions are only affected by the first input $u_{(1)}$, which is the same in both cases."]}
{"annotations": [{"id": 86979248, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86986389, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86988461, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-00514a3d604be5ad.jsonl", "id_parag": "2310.19083-00514a3d604be5ad_4", "final_parag": "Let us now address some critical aspects regarding our proposed backward reachability algorithms: First of all, the target set $\\targetset{}$ has to be represented as a polytope. \nWhile the manual design of polytopes is quite intuitive, the target set may come from another algorithm and thus be represented by a different set representation, which needs to be converted to a polytope. \nFor minimal reachability, one wants an outer approximation of the original set, whereas maximal reachability requires the converted polytope to be contained in the original set---both cases can be handled via optimization, e.g., using support function evaluations \\eqref{eq:sFset} to obtain an enclosing polytope.", "candidate_comments": ["Another future direction is to generalize the presented backward reachability algorithms to nonlinear continuous-time systems. \nA well-known issue in using set propagation for forward reachability of nonlinear systems is the wrapping effect, which can lead to an explosion of the set size over time. \nStill, some ideas outlined in this article may be reusable, such as \\cref{fig:BRSAE_ti} depicting the intersection of various backward reachable sets in order to iteratively exclude states from the minimal backward reachable set."]}
{"annotations": [{"id": 86986714, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86986867, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86989188, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-00514a3d604be5ad.jsonl", "id_parag": "2310.19083-00514a3d604be5ad_5", "final_parag": "As discussed in the respective subsections, the approximation errors of all backward reachable sets except the time-point maximal backward set are non-zero even in the limit $\\Delta t \\to 0$. \nStill, one can tighten the time-point and time-interval minimal backward reachable sets in arbitrary directions by additional support function evaluations. \nThese directions will be chosen according to the demands of the considered application scenario. \nFor the time-interval maximal backward reachable set, the approximation error entirely depends on the tightness of the containment in \\cref{prop:union}. \nFor large disturbances, the forward reachable set of a given initial state may not be contained within the target set at any specific point in time, but still pass through the target set over a time interval. \nThis would make that initial state part of the time-interval solution, while excluding it from the time-point solution. \nFurther investigation into this issue is required to formally capture the notion of one set passing through another, which is different from both containment and intersection.", "candidate_comments": ["Another future direction is to generalize the presented backward reachability algorithms to nonlinear continuous-time systems. \nA well-known issue in using set propagation for forward reachability of nonlinear systems is the wrapping effect, which can lead to an explosion of the set size over time. \nStill, some ideas outlined in this article may be reusable, such as \\cref{fig:BRSAE_ti} depicting the intersection of various backward reachable sets in order to iteratively exclude states from the minimal backward reachable set."]}
{"annotations": [{"id": 86904744, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973042, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86974731, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86988700, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-19e054f9fc31aef7.jsonl", "id_parag": "2310.19083-19e054f9fc31aef7_1", "final_parag": "We introduce some general notation, basics of set-based arithmetic, and fundamentals on forward reachability analysis required for the main body of this article.", "candidate_comments": ["We introduce the operations $\\min \\I{} = a$ and $\\max \\I{} = b$, returning the infimum and supremum, respectively."]}
{"annotations": [{"id": 86962515, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86980096, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86980859, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-19e054f9fc31aef7.jsonl", "id_parag": "2310.19083-19e054f9fc31aef7_2", "final_parag": "The set of real numbers is denoted by $\\R{}$, the set of natural numbers without zero is denoted by $\\N{}$, and the subset $\\{a,a+1,...,b\\} \\subset \\N{}$ for $0 < a < b$, is denoted by $\\Nint{a}{b}$. \nWe denote scalars and vectors by lowercase letters and matrices by uppercase letters. \nFor a vector $s \\in \\R{n}$, $\\norm{s}_p$ returns its $p$-norm and $s_{(i)}$ represents its $i$th entry; for a matrix $M \\in \\R{m \\times n}$, $M_{(i,\\cdot)}$ refers to the $i$th row and $M_{(\\cdot,j)}$ to the $j$th column. \nThe operation $\\diag(s)$ returns a square matrix with the vector~$s$ on its main diagonal. \nHorizontal concatenation of two properly-sized matrices $M_1$ and $M_2$ is denoted by $[M_1\\;M_2]$ and the identity matrix of dimension~$n$ by $I_n$. \nFurthermore, we use $\\matzeros{}$ and $\\vecones{}$ to represent vectors and matrices of proper dimension containing only zeros or ones. \nWe denote exact sets by standard calligraphic letters $\\S{}$, inner approximations by~$\\widecheck{\\mathcal{S}}$, and outer approximations by $\\widehat{\\mathcal{S}}$. \nWe write the set $\\{-s | s \\in \\S{} \\}$ as $-\\S{}$ and represent the empty set by $\\emptyset$.", "candidate_comments": ["We introduce the operations $\\min \\I{} = a$ and $\\max \\I{} = b$, returning the infimum and supremum, respectively."]}
{"annotations": [{"id": 86952985, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86959552, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86964845, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-19e054f9fc31aef7.jsonl", "id_parag": "2310.19083-19e054f9fc31aef7_3", "final_parag": "Interval matrices extend intervals by using matrices as lower and upper limits and are denoted in bold calligraphic letters, e.g. $\\intmat{I}$. \nThe operations $\\centerOp{\\S{}}$ and $\\boxOp{\\S{}}$ compute the volumetric center and tightest axis-aligned interval outer approximation of the set $\\S{}$, respectively.", "candidate_comments": ["We introduce the operations $\\min \\I{} = a$ and $\\max \\I{} = b$, returning the infimum and supremum, respectively."]}
{"annotations": [{"id": 86868677, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86900275, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86900784, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19083-19e054f9fc31aef7.jsonl", "id_parag": "2310.19083-19e054f9fc31aef7_4", "final_parag": "where $A \\in \\R{n \\times n}$ is the system matrix in \\cref{def:FRS} and the interval matrix $\\E{}$ is the remainder of the exponential matrix \\cite[Eq.~(3.2)]{Althoff2010diss}:", "candidate_comments": [", whereas the linear map in \\eqref{eq:innerZ} can be evaluated similarly to the common matrix exponential $e^{At}$ by extracting $A^{-1}$."]}
{"annotations": [{"id": 86723478, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86724473, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86761805, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19083-80278337999fe4ea.jsonl", "id_parag": "2310.19083-80278337999fe4ea_1", "final_parag": "Since the respective sets are closed under the applied operations, the entire approximation error is incurred by the outer and inner approximation of the particular solutions. \nSince these approximations converge to their exact counterparts in the limit $\\Delta t \\to 0$ by \\cref{prop:convergence}, the approximation errors of $\\outerBRSEA{-t}$ in \\eqref{eq:outerBRSEA_tp} and $\\innerBRSEA{-t}$ in \\eqref{eq:innerBRSEA_tp} also converge to $0$ as $\\Delta t \\to 0$.", "candidate_comments": ["which we can evaluate using \\eqref{eq:dH_Zprop} and \\eqref{eq:dH_Zinnerouter}. \nWe obtain the same formula as in \\eqref{eq:dH_BRSEA} for the approximation error of the outer approximation $d_H(\\BRSEA{-t},\\outerBRSEA{-t})$. \nNote that this error converges to 0 for $\\Delta t \\to 0$."]}
{"annotations": [{"id": 86733266, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86902027, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86903517, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-80278337999fe4ea.jsonl", "id_parag": "2310.19083-80278337999fe4ea_2", "final_parag": "\\begin{lemma}[Distributivity of Minkowski difference over convex hull] For three compact, convex, and nonempty sets $\\S{1}, \\S{2}, \\S{3} \\subset \\R{n}$, we have [EQUATION] \\end{lemma} \\begin{proof} See Appendix. \n\\end{proof}", "candidate_comments": ["\\begin{lemma}[Distributivity of Minkowski difference over linear map]  For an invertible matrix $M \\in \\R{n \\times n}$ and two compact, convex, and nonempty sets $\\S{1}, \\S{2} \\in \\R{n}$, we have"]}
{"annotations": [{"id": 86873432, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86901563, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86903757, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-80278337999fe4ea.jsonl", "id_parag": "2310.19083-80278337999fe4ea_3", "final_parag": "The union over all $\\steps{}$ steps, that is, [EQUATION] is an inner approximation of the time-interval backward reachable set $\\BRSEA{-\\tau}$ in \\eqref{eq:def_BRSEA_ti} over the time interval $\\tau = [t_0,\\tFinal{}]$. \n\\end{theorem} \\begin{proof} See Appendix. \n\\end{proof}", "candidate_comments": ["\\begin{corollary}  As the time step size goes to $0$, our inner approximation of the time-interval solution $\\innerBRSEA{-\\tau_k}$ converges to the inner approximation of the time-point solution $\\innerBRSEA{-t_k}$, i.e.,"]}
{"annotations": [{"id": 86985710, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86987068, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86988511, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-80278337999fe4ea.jsonl", "id_parag": "2310.19083-80278337999fe4ea_4", "final_parag": "\\cref{alg:BRSEA_ti} implements \\cref{thm:BRSEA_ti}, where we explicitly consider the more general case of a time interval $\\tau = [t_0,\\tFinal{}]$ with $t_0 > 0$: We pre-compute the particular solutions $\\innerZU{t}$ and $\\outerZW{t}$ until time $t_0$ in line~\\ref{alg:BRSEA_ti:initZUZW} and pre-compute the polytopes $\\poly{1}, \\poly{2}$ (lines~\\ref{alg:BRSEA_ti:P1}-\\ref{alg:BRSEA_ti:P2}) that are used for inner approximating the time-interval homogeneous solution, see \\eqref{eq:innerHti}. \nThe main loop computes all individual backward reachable sets $\\innerBRSEA{-\\tau_k}$ following \\cref{thm:BRSEA_ti}, whose union (\\cref{alg:BRSEA_ti:union}) yields the inner approximation of the time-interval maximal backward reachable set $\\innerBRSEA{-\\tau}$.", "candidate_comments": ["\\begin{corollary}  As the time step size goes to $0$, our inner approximation of the time-interval solution $\\innerBRSEA{-\\tau_k}$ converges to the inner approximation of the time-point solution $\\innerBRSEA{-t_k}$, i.e.,"]}
{"annotations": [{"id": 86982763, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86984461, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86984586, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19083-88f7da84fe68d42f.jsonl", "id_parag": "2310.19083-88f7da84fe68d42f_1", "final_parag": "Let us briefly consider autonomous systems $\\dot{x} = f(x)$, where the backward reachable set is equal to the forward reachable set for the time-inverted dynamics $\\dot{x} = -f(x)$ using the target set $\\targetset{}$ as the initial set. \nIf the target set represents an unsafe set, one can use established forward reachability algorithms for computing outer approximations of linear systems \\cite{Girard2006HSCC1,LeGuernic2010NAHS} and nonlinear systems \\cite{Althoff2008CDC,Chen2015diss}. \nIf on the other hand, the target set is a goal set, we require to compute an inner approximation, for which there also exist many methods for linear systems \\cite{Girard2006HSCC1,Frehse2015CDC} as well as nonlinear systems \\cite{Xue2017a,Goubault2017HSCC,Goubault2020CSL,Chen2014FMCAD,Kochdumper2020CDC}. \nSince this very special case is not the focus of our work, we do not discuss the different approaches here and instead refer the interested reader to the overviews in the cited literature.", "candidate_comments": ["If on the other hand, the target set is a goal set, we require to compute an inner approximation: Works for linear time-invariant systems use the convex hull of extremal points \\cite{Girard2006} or simulation runs \\cite{Frehse2015CDC} for time-point solutions. \nTime-interval solutions can be inner approximated by the Minkowski difference of a computed outer approximation with a set accounting for the approximation error \\cite[Sec.~5]{Wetzlinger2023TAC}."]}
{"annotations": [{"id": 87251432, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87258940, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87267911, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-88f7da84fe68d42f.jsonl", "id_parag": "2310.19083-88f7da84fe68d42f_2", "final_parag": "An approach for decoupled dynamics has been presented in \\cite{Chen2015CDC}. \nIn the context of systems coupled by multi-agent interaction, the decoupled computation has been augmented by a higher-level control using mixed integer programming \\cite{Chen2016CDC}. \nMoreover, a deep neural network has been trained to output the value function describing the reachable set, which improves the scalability but invalidates all safety guarantees \\cite{Bansal2021ICRA}. \nOther ideas to improve performance include warm-starting and adaptive grid sampling \\cite{Herbert2021ICRA}.", "candidate_comments": ["Furthermore, HJ reachability has been combined with reinforcement learning, enabling the analysis of up to $18$-dimensional systems at the cost of safety guarantees \\cite{Fisac2019}."]}
{"annotations": [{"id": 86971857, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86975112, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86977188, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19083-c40c44d3bd98be61.jsonl", "id_parag": "2310.19083-c40c44d3bd98be61_1", "final_parag": "If the target set represents an unsafe set, one utilizes the notion of \\emph{minimal} reachability \\cite[Sec.~4.2]{Mitchell2007HSCC}: The minimal backward reachable set contains all states that cannot avoid entering the target set regardless of the chosen control input. \nConsequently, all states within the backward reachable set are deemed unsafe and thus should also be avoided. \nIn case an exact solution cannot be obtained, we resort to computing outer approximations to maintain safety. \nA common example is obstacle avoidance: The target set represents the obstacle and the minimal backward reachable set contains all states from which one cannot avoid hitting the obstacle.", "candidate_comments": ["For instance, the maximal backward reachable set of a quadrotor for a target set contains all states from which the quadrotor can reach the target."]}
{"annotations": [{"id": 86977631, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86981968, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86981536, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-c40c44d3bd98be61.jsonl", "id_parag": "2310.19083-c40c44d3bd98be61_2", "final_parag": "If the target set represents a goal set, the concept of \\emph{maximal} reachability \\cite[Sec.~4.1]{Mitchell2007HSCC} is applicable: The maximal backward reachable set contains all states from which we can steer into the target set despite worst-case disturbances. \nNote that any initial state only requires to reach the target set by a single control input trajectory to become part of the backward reachable set. \nTo ensure that all contained initial states can definitely be steered into the target set, we require an inner approximation if the exact solution cannot be computed.", "candidate_comments": ["For instance, the maximal backward reachable set of a quadrotor for a target set contains all states from which the quadrotor can reach the target."]}
{"annotations": [{"id": 86904795, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86971274, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86971157, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19083-c40c44d3bd98be61.jsonl", "id_parag": "2310.19083-c40c44d3bd98be61_3", "final_parag": "Maximal backward reachability is closely related to controller synthesis: The backward reachable set contains all states for which a controller exists such that the target set is reachable.", "candidate_comments": ["For instance, the maximal backward reachable set of a quadrotor for a target set contains all states from which the quadrotor can reach the target."]}
{"annotations": [{"id": 86961976, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86973378, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}, {"id": 86973836, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19083-c40c44d3bd98be61.jsonl", "id_parag": "2310.19083-c40c44d3bd98be61_4", "final_parag": "In this article, we compute minimal and maximal backward reachable sets for continuous-time linear time-invariant (LTI) systems. \nAs there are many similar definitions of backward reachable sets as well as related concepts, we postpone the literature review to \\cref{sec:relatedwork}. \nThis allows us to use the preliminary information from \\cref{sec:preliminaries,sec:problemstatement} for a more concise overview. \nOur contributions are as follows:", "candidate_comments": ["For instance, the maximal backward reachable set of a quadrotor for a target set contains all states from which the quadrotor can reach the target.", "\\item An extension of the proposed maximal backward reachable set computation to respect state constraints over the entire time horizon (\\cref{ssec:BRSEA_constr})."]}
{"annotations": [{"id": 86960033, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86976426, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}, {"id": 86978137, "completed_by": {"id": 140156}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19083-c40c44d3bd98be61.jsonl", "id_parag": "2310.19083-c40c44d3bd98be61_5", "final_parag": "\\begin{itemize} \\item An inner and outer approximation for the time-point minimal backward reachable set (\\cref{ssec:BRSAE_tp}). \n\\item An outer approximation of the time-interval minimal backward reachable set (\\cref{ssec:outerBRSAE_ti}). \n\\item An inner and outer approximation for the time-point maximal backward reachable set (\\cref{ssec:BRSEA_tp}). \n\\item An inner approximation for the time-interval maximal backward reachable set (\\cref{ssec:BRSEA_ti}).", "candidate_comments": ["For instance, the maximal backward reachable set of a quadrotor for a target set contains all states from which the quadrotor can reach the target.", "\\item An extension of the proposed maximal backward reachable set computation to respect state constraints over the entire time horizon (\\cref{ssec:BRSEA_constr})."]}
{"annotations": [{"id": 86864491, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86902003, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86903538, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-cda0fcec3a49cc6f.jsonl", "id_parag": "2310.19083-cda0fcec3a49cc6f_1", "final_parag": "In general, backward reachability analysis aims to compute the set of states that reach a target set $\\targetset{} \\subset \\R{n}$ after a certain elapsed time $t$ (time-point backward reachable set) or at any time within the interval $\\tau = [t_0,\\tFinal{}]$ (time-interval backward reachable set). \nWe assume the target set $\\targetset{} \\subset \\R{n}$ to be represented as a polytope. \nLet us first define the \\emph{minimal} backward reachable set, where the target set is composed of unsafe states:", "candidate_comments": ["We are mainly interested in computing outer approximations of the minimal backward reachable sets \\eqref{eq:def_BRSAE_tp}-\\eqref{eq:def_BRSAE_ti} to enclose all states leading to unsafe system behavior."]}
{"annotations": [{"id": 86735471, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86899723, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86900691, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-cda0fcec3a49cc6f.jsonl", "id_parag": "2310.19083-cda0fcec3a49cc6f_2", "final_parag": "Case \\ding{192} in \\cref{fig:BRS} illustrates the time-point set \\eqref{eq:def_BRSAE_tp}: For all states within the minimal backward reachable set $\\BRSAE{-t}$, such as $x_0^{(1)}$, the target set $\\targetset{}$ is unavoidable regardless of the input signal $u^{(1)}(\\cdot)$. \nFor any initial state outside $\\BRSAE{-t}$ like $x_0^{(2)}$, there is at least one input signal $u^{(2)}(\\cdot)$, for which there is no disturbance signal such that the corresponding forward reachable set intersects $\\targetset{}$.", "candidate_comments": ["We are mainly interested in computing outer approximations of the minimal backward reachable sets \\eqref{eq:def_BRSAE_tp}-\\eqref{eq:def_BRSAE_ti} to enclose all states leading to unsafe system behavior."]}
{"annotations": [{"id": 86737479, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86899624, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86900523, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-cda0fcec3a49cc6f.jsonl", "id_parag": "2310.19083-cda0fcec3a49cc6f_3", "final_parag": "In the following definition of the \\emph{maximal} backward reachable set, the target set represents a goal set into which we want to steer the state despite worst-case disturbances.", "candidate_comments": ["We are mainly interested in computing outer approximations of the minimal backward reachable sets \\eqref{eq:def_BRSAE_tp}-\\eqref{eq:def_BRSAE_ti} to enclose all states leading to unsafe system behavior."]}
{"annotations": [{"id": 86725993, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86731303, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86826323, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-cda0fcec3a49cc6f.jsonl", "id_parag": "2310.19083-cda0fcec3a49cc6f_4", "final_parag": "The time-interval maximal backward reachable set \\cite[Def.~3]{Chen2018ANNUREV} requires the state to pass through $\\targetset{}$ anytime in the time interval $\\tau$.", "candidate_comments": ["We want to compute inner approximations of maximal backward reachable sets \\eqref{eq:def_BRSEA_tp}-\\eqref{eq:def_BRSEA_ti} such that the contained states are guaranteed to reach $\\targetset{}$ in the presence of worst-case disturbances."]}
{"annotations": [{"id": 86734490, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86772256, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86826780, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-dd0212a15f4f4aea.jsonl", "id_parag": "2310.19083-dd0212a15f4f4aea_1", "final_parag": "where we introduce the operator $\\ooplus$ to distinguish this operation from the exact Minkowski sum. \nThe runtime complexity is $\\bigO{\\cons{} n\\gamma}$. \n\\end{proposition} \\begin{proof} See Appendix. \n\\end{proof}", "candidate_comments": ["We will also evaluate the $2n$ axis-aligned support functions, which increases the runtime complexity to $\\bigO{(\\cons{}+2n)n\\gamma}$."]}
{"annotations": [{"id": 86962190, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86972505, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973642, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19083-dd0212a15f4f4aea.jsonl", "id_parag": "2310.19083-dd0212a15f4f4aea_2", "final_parag": "Under \\cref{ass:runtimecomplexity} and following \\cref{tab:setops}, the outer approximative Minkowski sum from \\cref{prop:minkSum_polyzono}, the Minkowski difference, and the linear map in the computation of the outer approximation $\\outerBRSAE{-t}$ are all $\\bigO{n^3}$, while the computation of the inner approximation $\\innerBRSAE{-t}$ is dominated by the conversion to a constrained zonotope, which is $\\bigO{n^4}$.", "candidate_comments": ["Under \\cref{ass:runtimecomplexity}, the computation of the outer approximation $\\outerBRSAE{-t}$ is marginally dominated by the over-approximative Minkowski sum, which is $\\bigO{(\\cons{}+2n)n, since the Minkowski difference and linear map are at most $\\bigO{(\\cons{}+2n)n \\steps{}n}$ and $\\bigO{(\\cons{}+2n)n^2}$, respectively; all these operations are essentially $\\bigO{n^3}$ for $n \\gg \\steps{}$ and under \\cref{ass:runtimecomplexity}."]}
{"annotations": [{"id": 86865240, "completed_by": {"id": 70661}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86902456, "completed_by": {"id": 62471}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86901929, "completed_by": {"id": 62306}, "result": ["Différent", "Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19083-f52570d8d336f342.jsonl", "id_parag": "2310.19083-f52570d8d336f342_1", "final_parag": "\\emph{Proof of \\cref{prop:minkSum_polyzono}}: \\\\ We insert $\\poly{} \\oplus \\Z{}$ into \\eqref{eq:sFset} to obtain The runtime complexity follows from the $\\cons{}$ support function evaluations of $\\Z{}$ \\eqref{eq:sF_zono}. \n\\hfill $\\square$", "candidate_comments": ["\\emph{Proof of \\cref{prop:union}}: \\\\ The reason is the order of quantifiers \\cite[Prop.~2]{Mitchell2007}. \n\\hfill $\\square$", "\\emph{Proof of \\cref{prop:BRSAE_tp}}: \\\\ This is a continuization of the discrete-time case proven in \\cite[Thm.~2.4]{Kurzhanskiy2011}. \n\\hfill $\\square$", "Note that we convert the polytopic target set $\\targetset{}$ to a constrained zonotope using \\eqref{eq:conZonoOp} for an efficient evaluation of the convex hull operation. \n\\hfill $\\square$"]}
{"annotations": [{"id": 86961607, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972330, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86972358, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-f52570d8d336f342.jsonl", "id_parag": "2310.19083-f52570d8d336f342_2", "final_parag": "\\emph{Proof of \\cref{thm:BRSAE_ti}}: \\\\ By considering only a finite subset of input signals $\\inputsignalssub{} \\subset \\inputsignals{}$, we obtain an outer approximation:", "candidate_comments": ["\\emph{Proof of \\cref{prop:BRSAE_tp}}: \\\\ This is a continuization of the discrete-time case proven in \\cite[Thm.~2.4]{Kurzhanskiy2011}. \n\\hfill $\\square$"]}
{"annotations": [{"id": 86986486, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}, {"id": 86986967, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}, {"id": 86989231, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19083-f52570d8d336f342.jsonl", "id_parag": "2310.19083-f52570d8d336f342_3", "final_parag": "\\emph{Proof of \\cref{lmm:convMinkDiff}}: \\\\ We plug into the definitions of the Minkowski difference \\eqref{eq:def_minkDiff} and convex hull \\eqref{eq:def_conv}: from which it follows that [EQUATION] since $\\S{} \\oplus \\S{3} \\ominus \\S{3} = \\S{}$ holds \\cite[Lemma~1(iii)]{Yang2022CSL}. \n\\hfill $\\square$", "candidate_comments": ["\\emph{Proof of \\cref{prop:BRSEA_tp}}: \\\\ This is a continuization of the discrete-time case proven in \\cite[Thm.~2.4]{Kurzhanskiy2011}. \n\\hfill $\\square$", "currently not needed \\emph{Proof of \\cref{lmm:linMapMinkDiff}}: \\\\ Using the matrix inverse $M^{-1} \\in \\R{n \\times n}$, we have which is equal to $M(\\S{1} \\ominus \\S{2}) = \\{ Ms | s \\oplus \\S{2} \\subseteq \\S{1} \\}$ by variable transform. \n\\hfill $\\blacksquare$"]}
{"annotations": [{"id": 86721962, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86730802, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86729966, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-f52570d8d336f342.jsonl", "id_parag": "2310.19083-f52570d8d336f342_4", "final_parag": "\\emph{Proof of \\cref{thm:BRSEA_ti}}: \\\\ A single time-interval solution $\\BRSEA{-\\tau_k}$ over $\\tau_k = [t_k,t_{k+1}]$ covering part of the union in \\eqref{eq:BRSEA_tp_union} can be expressed by", "candidate_comments": ["\\emph{Proof of \\cref{prop:BRSEA_tp}}: \\\\ This is a continuization of the discrete-time case proven in \\cite[Thm.~2.4]{Kurzhanskiy2011}. \n\\hfill $\\square$"]}
{"annotations": [{"id": 86960950, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972939, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86973924, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19083-f52570d8d336f342.jsonl", "id_parag": "2310.19083-f52570d8d336f342_5", "final_parag": "\\caption{Generator matrix $G$ of the safe terminal set $\\zono{\\matzeros{},G}$ for the quadrotor system in \\cref{ssec:terminalset} computed using the approach in \\cite{Gruber2021CSL}.}", "candidate_comments": ["\\caption{Dynamics equations with $g = 9.81$ for the form \\eqref{eq:linsys} for the quadrotor system analyzed in \\cref{ssec:terminalset}.}"]}
{"annotations": [{"id": 86983269, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985432, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86985624, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19084-f7e6a5f92dad7d5f.jsonl", "id_parag": "2310.19084-f7e6a5f92dad7d5f_1", "final_parag": "\\item Higher human resemblance is significantly correlated to better language modeling. Scaling improves the resemblance by scaling law \\citep{henighan_scaling_2020}, while instruction tuning reduces it. All models have higher resemblance to L2 rather than to L1, suggesting further room for the improvement in language perception (\\S \\ref{ss:human_result}).", "candidate_comments": ["\\item Instruction tuning has limited effect on the general attention distribution on plain text, but enhance the model's sensitivity to instructions. It reduces the human resemblance, increases the trivial pattern reliance. "]}
{"annotations": [{"id": 86962710, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86971429, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86971487, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19084-f7e6a5f92dad7d5f.jsonl", "id_parag": "2310.19084-f7e6a5f92dad7d5f_2", "final_parag": "\\paragraph{Instruction tuned LLMs} achieve better performances in performing tasks \\citep{ouyang_training_2022,wei_finetuned_2022,sanh_multitask_2022}. It seems that the LLMs are better at understanding human instructions. However, there is little discussion on how the instruction tuning process affects the language perception.", "candidate_comments": ["\\item Instruction tuning has limited effect on the general attention distribution on plain text, but enhance the model's sensitivity to instructions. It reduces the human resemblance, increases the trivial pattern reliance. "]}
{"annotations": [{"id": 86961055, "completed_by": {"id": 70661}, "result": ["Différent", "Différent", "Réécriture"]}, {"id": 86959957, "completed_by": {"id": 62471}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86910557, "completed_by": {"id": 140156}, "result": ["Différent", "Réécriture", "Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19084-f7e6a5f92dad7d5f.jsonl", "id_parag": "2310.19084-f7e6a5f92dad7d5f_3", "final_parag": "To show the effect of scaling and instruction tuning on different models, we compare the self-attention scores of different LLMs given the same input, by viewing model attention as probability distributions and calculating the general attention divergence based on Jensen-Shannon divergence.", "candidate_comments": ["To clarify, the focus of this work is not interpreting the model prediction by attention, but is analyzing the attention patterns with the help of human attention.", "The basic idea is to compare the self-attention of models in different scales and training stages to study the effects of the two factors on language perception. To do this, we need to answer two questions: \\begin{enumerate}     \\item Do the two factors in general cause significant change in model attention?     \\item Do scaling or instruction tuning makes model attention better or worse for language perception? \\end{enumerate}", "To answer the first question, we need to compare the self-attention scores of different models. Here, we view model attention as probability distributions, and use the divergence between distributions to calculate their difference."]}
{"annotations": [{"id": 86986174, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86986899, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86986931, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19084-f7e6a5f92dad7d5f.jsonl", "id_parag": "2310.19084-f7e6a5f92dad7d5f_4", "final_parag": "In addition, we select several trivial, context-free attention patterns, and design a trivial pattern reliance metric to help demonstrate the difference between models and human subject groups.", "candidate_comments": ["To answer the first question, we need to compare the self-attention scores of different models. Here, we view model attention as probability distributions, and use the divergence between distributions to calculate their difference."]}
{"annotations": [{"id": 86981643, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985772, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86985806, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19084-f7e6a5f92dad7d5f.jsonl", "id_parag": "2310.19084-f7e6a5f92dad7d5f_5", "final_parag": "We assesses the mean Jensen-Shannon (J-S) divergence between attention from different models on all sentences. For two attention matrices $A,\\ B\\in\\mathbb R^{n_{ \\mathrm{word}}\\times n_{\\mathrm{word}}}$, the J-S divergence is:", "candidate_comments": ["To answer the first question, we need to compare the self-attention scores of different models. Here, we view model attention as probability distributions, and use the divergence between distributions to calculate their difference."]}
{"annotations": [{"id": 86952429, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86959136, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86970595, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19084-f7e6a5f92dad7d5f.jsonl", "id_parag": "2310.19084-f7e6a5f92dad7d5f_6", "final_parag": "To identify high and low divergence, a reference value is required. We propose to use the attention divergence between two models that share the same structure, size, and training data, but have small difference in training strategies as the reference. A divergence value higher than the reference suggests relatively large change attention, while one lower than that indicates small (but not necessarily negligible) change.", "candidate_comments": ["The comparison has three steps. First, we compare the general attention distribution of different models by calculating the divergence between them. Next, we define three specific measurements, namely human resemblance, trivial pattern reliance, and sensitivity to instructions, and compare these values of different models. Last, to explain the relation between our main measurement, i.e. human resemblance, and model performance, we calculate the correlation between them and analyze the human attention pattern of different groups."]}
{"annotations": [{"id": 86980181, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86986356, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86986392, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19084-f7e6a5f92dad7d5f.jsonl", "id_parag": "2310.19084-f7e6a5f92dad7d5f_7", "final_parag": "\\subsubsection{Scaling Reduces Trivial Pattern Reliance}", "candidate_comments": ["This shows scaling effectively reduces the models' trivial pattern reliance, especially in the deeper layers"]}
{"annotations": [{"id": 86960519, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972402, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86972467, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19096-00421b8754e33bff.jsonl", "id_parag": "2310.19096-00421b8754e33bff_1", "final_parag": "\\begin{definition}[Realization] We define $\\realizes{\\valOne}{\\lcOne}{\\typeOne}$, which reads \\emph{$\\valOne$ realizes $\\typeOne$ under $\\lcOne$}, as the smallest relation such that \\end{definition}", "candidate_comments": ["\\begin{definition}[Realization] We define $\\realizes{\\valOne}{\\lcOne}{\\typeOne}$ as the smallest relation such that If $\\realizes{\\valOne}{\\lcOne}{\\typeOne}$ we say that \\emph{$\\valOne$ realizes $\\typeOne$ under $\\lcOne$}. \n\\end{definition}"]}
{"annotations": [{"id": 86981768, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86986264, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86986415, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19096-00421b8754e33bff.jsonl", "id_parag": "2310.19096-00421b8754e33bff_2", "final_parag": "By themselves, realization and reducibility are defined only on terms and values closed in the sense both of regular and index variables. To extend these notions to open terms and values, we adopt the standard approach of reasoning explicitly about the substitutions that could render them closed. \n\\begin{definition}[Closing Substitution] We define the set $\\vsubSet$ of \\emph{closing value substitutions} as the smallest subset of $\\valSet\\cup\\termSet \\to \\valSet\\cup\\termSet$ such that \\begin{itemize} \\item $\\emptysub\\in\\vsubSet$ with $\\emptysub(\\termOne) = \\termOne$. \n\\item If $\\vsubOne\\in\\vsubSet$, $\\varOne$ is a variable name and $\\valOne\\in\\valSet$ is closed, then $\\vsubOne[\\varOne\\mapsto\\valOne]\\in\\vsubSet$ with $\\vsubOne[\\varOne\\mapsto\\valOne](\\termOne) = \\vsubOne(\\termOne\\sub{\\valOne}{\\varOne})$. \n\\end{itemize} We define the set $\\isubSet$ of \\emph{closing index substitutions} as the smallest subset of $\\indexSet\\cup\\typeSet\\cup\\valSet\\cup\\termSet \\to \\indexSet\\cup\\typeSet\\cup\\valSet\\cup\\termSet$ such that \\begin{itemize} \\item $\\emptysub\\in\\isubSet$ with $\\emptysub(\\termOne) = \\termOne$. \n\\item If $\\isubOne\\in\\isubSet$, $\\ivarOne$ is an index variable name and $\\indexOne\\in\\indexSet$ is closed, then $\\isubOne[\\ivarOne\\mapsto\\indexOne]\\in\\isubSet$ with $\\isubOne[\\ivarOne\\mapsto\\indexOne](\\termOne) = \\isubOne(\\termOne\\isub{\\indexOne}{\\ivarOne})$. \n\\end{itemize} \\end{definition}", "candidate_comments": ["\\begin{definition}[Realization] We define $\\realizes{\\valOne}{\\lcOne}{\\typeOne}$ as the smallest relation such that If $\\realizes{\\valOne}{\\lcOne}{\\typeOne}$ we say that \\emph{$\\valOne$ realizes $\\typeOne$ under $\\lcOne$}. \n\\end{definition}"]}
{"annotations": [{"id": 86961883, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972285, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86972308, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_1", "final_parag": "Jiayao Tan\\textsuperscript{\\rm 1}\\equalcontrib, Fan Lyu\\textsuperscript{\\rm 2}\\equalcontrib, Linyan Li\\textsuperscript{\\rm 3}, Fuyuan Hu\\textsuperscript{\\rm 1}\\thanks{Corresponding author.}, Tingliang Feng\\textsuperscript{\\rm 4}, Fenglei Xu\\textsuperscript{\\rm 1}, Rui Yao } \\affiliations{", "candidate_comments": ["Sunil Issar\\textsuperscript{\\rm 2},  J. Scott Penberthy\\textsuperscript{\\rm 3},  George Ferguson\\textsuperscript{\\rm 4}, Hans Guesgen\\textsuperscript{\\rm 5} Note that the comma should be placed after the superscript"]}
{"annotations": [{"id": 86865009, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86901802, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86902181, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_2", "final_parag": "\\textsuperscript{\\rm 1}Suzhou University of Science and Technology, \\textsuperscript{\\rm 2}CRIPAC, MAIS, CASIA,\\\\ \\textsuperscript{\\rm 3}Suzhou Institute of Trade Commerce, \\textsuperscript{\\rm 4}Tianjin University\\\\ \\{jiayaotan@post,\\textsuperscript{\\rm †}fuyuanhu@mail,xufl@mail\\}.usts.edu.cn, fan.lyu@cripac.ia.ac.cn, lilinyan@szjm.edu.cn", "candidate_comments": ["Sunil Issar\\textsuperscript{\\rm 2},  J. Scott Penberthy\\textsuperscript{\\rm 3},  George Ferguson\\textsuperscript{\\rm 4}, Hans Guesgen\\textsuperscript{\\rm 5} Note that the comma should be placed after the superscript"]}
{"annotations": [{"id": 86983206, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86981710, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86984256, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_3", "final_parag": "Vehicle-to-everything (V2X) perception is an innovative technology that enhances vehicle perception accuracy, thereby elevating the security and reliability of autonomous systems. \nHowever, existing V2X perception methods focus on static scenes from mainly vehicle-based vision, which is constrained by sensor capabilities and communication loads. \nTo adapt V2X perception models to dynamic scenes, we propose to build V2X perception from road-to-vehicle vision and present \\emph{Adaptive Road-to-Vehicle Perception} (AR2VP) method.", "candidate_comments": ["Vehicle-to-everything (V2X) perception is a novel technology of providing more accurate perception to vehicles, improving the security and reliability of autonomous."]}
{"annotations": [{"id": 86903797, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86951533, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86958519, "completed_by": {"id": 126844}, "result": ["Différent", "Réécriture"]}], "agreement": 55.55555555555555, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_4", "final_parag": "In AR2VP, we leverage roadside units to offer stable, wide-range sensing capabilities and serve as communication hubs. \nAR2VP is devised to tackle both intra-scene and inter-scene changes. For the former, we construct a dynamic perception representing module, which efficiently integrates vehicle perceptions, enabling vehicles to capture a more comprehensive range of dynamic factors within the scene.", "candidate_comments": ["Vehicle-to-everything (V2X) perception is a novel technology of providing more accurate perception to vehicles, improving the security and reliability of autonomous.", "In AR2VP, we have roadside unit provide stable and large-range sensing and serve as a communication center."]}
{"annotations": [{"id": 86828035, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86856708, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86898324, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_5", "final_parag": "Moreover, we introduce a road-to-vehicle perception compensating module, aimed at preserving the maximized roadside unit perception information in the presence of intra-scene changes. For inter-scene changes, we implement an experience replay mechanism leveraging the roadside unit's storage capacity to retain a subset of historical scene data, maintaining model robustness in response to inter-scene shifts. \nWe conduct perception experiment on 3D object detection and segmentation, and the results show that AR2VP excels in both performance-bandwidth trade-offs and adaptability within dynamic environments. Our code is available at: https://github.com/tjy1423317192/AP2VP", "candidate_comments": ["the results demonstrate that our AR2VP method not only achieves superior performance-bandwidth trade-offs but also exhibits improved adaptability to dynamic environments."]}
{"annotations": [{"id": 86975665, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986476, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86986460, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_6", "final_parag": "The Vehicle-to-everything (V2X) technique \\citep{63,8,9}, facilitating collaboration between vehicles and various other entities \\cite{13,14}, is introduced as a popular means to enhance the perception system for intelligent driving \\cite{2,AAAI4：dri}. \\emph{However, existing V2X research predominantly centers around static data (barely no entity or scene change), which inadequately addresses the safety prerequisites of vehicles in dynamic traffic environments.}", "candidate_comments": ["V2X technology \\cite{63,8,9} plays a vital role in communication, facilitating information exchange between vehicles and RSU, leading to a more comprehensive traffic environment perception. This collaboration among vehicles and RSU \\cite{13,14} enables diverse perception perspectives, extending the range to cover larger, long-distance areas, and previously inaccessible blocked regions. Consequently, it provides accurate and comprehensive traffic information, greatly supporting intelligent driving \\cite{2} decision-making and enhancing driving safety. \nVehicle-to-everything (V2X) technique \\citep{63,8,9}, enabling the collaboration between vehicles and many other entities \\cite{13,14}, is proposed to improve the perception system for intelligent driving \\cite{2}.  \\emph{However, the existing V2X studies focus on static data, which can hardly meet the real-world safety requirements of vehicles in dynamic traffic environments.} "]}
{"annotations": [{"id": 86984529, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86986020, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86988641, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_7", "final_parag": "Motivated by this, this paper proposes \\textbf{Adaptive Road-to-Vehicle Perception (AR2VP)} approach, which constructs a road-to-vehicle cooperative perception model tailored for dynamic environments. \nFirst, to effectively handle intra-scene changes, we design a dynamic perception representing module and road-to-vehicle perception compensating module.", "candidate_comments": ["which constructs an adaptable collaborative graph for intra-scene changes, improving the overall adaptability of vehicles to dynamic environments. "]}
{"annotations": [{"id": 86734446, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}, {"id": 86901464, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86901554, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}], "agreement": 55.55555555555556, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_8", "final_parag": "These modules collaboratively construct an adaptable graph catering to intra-scene changes, thereby enhancing vehicles' overall adaptability within dynamic scenarios. \nSecond, to effectively adapt to inter-scene changes, we present RSU experience replay, utilizing RSU storage capacity and integrating experience replay \\cite{AAAI1：conti,60,61} techniques in continual learning \\cite{58,58.1,58.2,58.3,59}, enabling vehicles to adapt to large-scale scene transitions beyond intra-scene changes. Our approach is validated on the tasks of scene segmentation and 3D object detection. \nThe results on V2X-Sim dataset~\\cite{63} show good perception performance and the adaptability of AR2VP to dynamic scenes.", "candidate_comments": ["which constructs an adaptable collaborative graph for intra-scene changes, improving the overall adaptability of vehicles to dynamic environments. ", "which use the perception from road side, enhancing the overall adaptability of vehicles to dynamic environments."]}
{"annotations": [{"id": 86962334, "completed_by": {"id": 62471}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86973331, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent", "Différent"]}, {"id": 86972591, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture", "Réécriture"]}], "agreement": 50.0, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_9", "final_parag": "\\begin{itemize} \\item To the best of our knowledge, we are the first to investigate the adaptability of V2X to dynamic scenes. Accordingly, we also proposes AR2VP approach for dynamic V2X perception. \n\\item To address intra-scene changes, we design Dynamic Perception Representing module and Road-to-Vehicle Perception Compensating module. These modules tap into the perceptual insights from the road side, thereby bolstering the overall adaptability of vehicles within dynamic scenes.", "candidate_comments": ["which constructs an adaptable collaborative graph for intra-scene changes, improving the overall adaptability of vehicles to dynamic environments. ", "which use the perception from road side, enhancing the overall adaptability of vehicles to dynamic environments.", "To address intra-scene changes, we propose RSU Experience Replay, enabling vehicles to adapt to large-scale scene transitions beyond intra-scene changes. \n我们是首个研究V2X对于动态场景适应性的，并提出了Adaptive Road-to-Vehicle Perception (AR2VP) network 对于intra—sence changes，我们设计了Dynamic Perception Represent module and Road-to-Vehicle Perception Compensate module，improving the overall adaptability of vehicles to dynamic environments。 对于intre—sence changes，我们提出了RSU Experience Replay， enabling vehicles to adapt to large-scale scene transitions beyond intra-scene changes."]}
{"annotations": [{"id": 86904629, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}, {"id": 86970941, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}, {"id": 86965475, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_10", "final_parag": "We observed that AR2VP demonstrates superior entity perception outcomes, achieving the highest overall perception performance. This analysis underscores that current V2X technologies rarely rely on RSUs to expand perception horizons. In contrast, AR2VP harnesses the latent strengths of RSUs to address intra-scene changes, which enhances the vehicle's ability to adapt to dynamic scenes, consequently elevating the overall perception capabilities. However, AR2VP does exhibit a performance drawback in pedestrian detection, implying a particular challenge in detecting small targets.", "candidate_comments": ["We find that AR2VP shows superior entity perception results, with overall perception performance being the best. This analysis suggests that existing V2X technologies merely utilize RSU to extend perception horizons. In contrast, AR2VP leverages the latent advantages of RSU to model intra-scene changes, further enhancing vehicle adaptability to scene dynamics, thereby augmenting overall perception capabilities. However, in comparison to V2V, AR2VP exhibits a performance disadvantage in pedestrian detection, indicating a certain discrepancy in detecting small targets.", "We see that AR2VP demonstrates superior overall perception performance.  Through analysis, we conclude that existing V2X technologies do not consider the impact of inter-scene changes on perception. In contrast, AR2VP fully leverages the storage capacity of RSU and introduces continuous learning knowledge to model inter-scene changes. This allows vehicles to adapt to new scenes while also reducing the degree of forgetting of previous scenes, demonstrating a high adaptability to inter-scene changes in perception and thereby enhancing the overall robustness of perception."]}
{"annotations": [{"id": 86727434, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86959980, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86963582, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_11", "final_parag": "AR2VP presents itself as a frontrunner in terms of overall perception performance. Upon analysis, it's evident that traditional V2X technologies disregard the influence of inter-scene changes on perception. In contrast, AR2VP optimally exploits the storage capacity of RSU and integrates continuous learning principles to effectively address inter-scene changes. This strategic approach empowers vehicles to assimilate new scenes while minimizing the extent of memory loss from prior scenes. This capability shows a strong adaptability to inter-scene changes in perception, thereby enhancing the global robustness of perception.", "candidate_comments": ["We see that AR2VP demonstrates superior overall perception performance.  Through analysis, we conclude that existing V2X technologies do not consider the impact of inter-scene changes on perception. In contrast, AR2VP fully leverages the storage capacity of RSU and introduces continuous learning knowledge to model inter-scene changes. This allows vehicles to adapt to new scenes while also reducing the degree of forgetting of previous scenes, demonstrating a high adaptability to inter-scene changes in perception and thereby enhancing the overall robustness of perception."]}
{"annotations": [{"id": 86737685, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86959151, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86961094, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_12", "final_parag": "Although RSU-ER can be applied to other models to mitigate forgetting, AR2VP notably demonstrates the most favorable suitability. Fig.~\\ref{Fig.8} portrays the learning of new scene data based on the old model, revealing the degree of memory forgetting from previous scenes. The observation is clear: V2V and Disco struggle to accommodate inter-scene changes, leading to significant memory loss from previous scenes. In contrast, AR2VP adeptly navigates inter-scene changes, exhibiting a higher retention of memories from prior scenes. This analysis underscores AR2VP's capacity for the lowest forgetting rate and the most proficient performance in addressing inter-scene changes.", "candidate_comments": ["通过分析可以得出，现有的V2X技术未考虑场景间变化对感知的影响。然而，AR2VP则充分利用了RSU的存储能力，并引入了连续学习的知识，以对场景间变化进行建模。这使得车辆在适应新场景的同时也减轻了对以前场景的遗忘程度，在感知方面表现出对场景间变化的高适应性，从而提高了整体感知的稳健性。 RSU Experience Replay (RSU-ER) also be applied to other models to achieve the suppression of forgetting, but AR2VP demonstrates the best applicability. \nFig.~\\ref{Fig.8} shows that the model learns new scene data based on the old model, showing the extent of forgetting of memories from previous scenes. We observe that V2V and Disco fail to adapt to inter-scene changes, resulting in catastrophic forgetting of memories from previous scenes. In contrast, AR2VP demonstrates the ability to adapt to inter-scene changes and maintains better retention of memories from previous scenes. Through analysis, AR2VP exhibits the lowest forgetting rate and performs the best in inter-scene changes."]}
{"annotations": [{"id": 86983419, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985732, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86985872, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_13", "final_parag": "1) Comparing to AR2VP, AR2VP/32 degrades by 1.31\\% in segmentation, still outperforming Disco in terms of mIoU. This means the compressing features in AR2VP does not significantly compromise perception performance.", "candidate_comments": ["AR2P可以利用autoencoder对特征进行压缩，减少特征传输所使用的通信带宽，Comparing to the AR2VP, AR2VP/32 degrade by 1.31\\% in segmentation.我们可以看出，对AR2P进行特征压缩并不会显著减低感知性能。 2) AR2VP/32 performs 32 times feature compression using an autoencoder, still outperforms Disco in terms of mIoU. \n3) Feature compression does not significantly affect seg. performance."]}
{"annotations": [{"id": 86723197, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86724110, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86724324, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19113-208a98625067cee4.jsonl", "id_parag": "2310.19113-208a98625067cee4_14", "final_parag": "In the future, based on our experimental findings, we intend to enhance the AR2VP's capability in recognizing small objects. This will involve further refinement of the model to ensure more accurate and effective identification of small entities.", "candidate_comments": ["In the future, based on our experiment, we plan to improve the ability of AR2VP on the recognition of small objects."]}
{"annotations": [{"id": 86904533, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86986433, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86989644, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19125-66c30bd111f2a070.jsonl", "id_parag": "2310.19125-66c30bd111f2a070_1", "final_parag": "The rest of this paper is structured according to guidelines from~\\cite{Wohlin12} on how to structure experiments in software engineering. As per their requirements, we first present our {\\bf Methods} which divide into: \\bi \\item The algorithms we apply to data (in \\S\\ref{isbse}, \\S\\ref{smo} and \\S\\ref{isbse}) and \\item The case studies used for those algorithms (see \\S\\ref{casestudies}); and \\item The protocol we used for human experiments (in \\S\\ref{rq1})\\footnote{ Our human investigations were carried with approval of the NC State  with the Investigator Review Board, IRB protocol \\#24233.}. \n\\ei The results from those methods are reported in our {\\bf Results} section, which is followed by a {\\bf Threats to Validity} section.", "candidate_comments": ["The rest of this paper is structured as follows.Section~\\S\\ref{background} presents the background leading to the development of this work as well as details the inner workings of the {\\IT} tool. Section~\\S\\ref{methods} offers details on the case studies and algorithms used in our experiments. Section~\\S\\ref{results} presents our results and answers our two research questions. Finally, the last sections present our analysis on the threats to this work, conclusion, and future work."]}
{"annotations": [{"id": 86962363, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86975586, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86975308, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19125-66c30bd111f2a070.jsonl", "id_parag": "2310.19125-66c30bd111f2a070_2", "final_parag": "The code for FLASH, written in Python, comes from the Nair et al. repository\\footnote{ \\url{https://github.com/FlashRepo}}. \nAraújo et al.  did not  offer a reproduction package for their work so we reimplemented their code from  their descriptions.", "candidate_comments": ["The code for that NGA is available at our Github site~\\footnote{\\url{https://github.com/zxcv123456qwe/iSNEAK/blob/main/NGA.ipynb}}."]}
{"annotations": [{"id": 86725579, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86725844, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86729771, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19125-66c30bd111f2a070.jsonl", "id_parag": "2310.19125-66c30bd111f2a070_3", "final_parag": "The models {\\em POM3A, POM3B, POM3C, OSP2, FLIGHT and GROUND} are small and have no constraints (so all solutions generated to these models are valid). \nAll the other models are much larger and have many constraints that only  1\\% to 3\\% of the solution space is valid.", "candidate_comments": ["For example, the SCRUM model of Fig. \\ref{fig:scrumModel} which is an SPL defining the possible legal configurations of the SCRUM organization framework. As shown on  this figure, the SCRUM model contains 128 features and more than 250 constraints. These constraints are so complicated that out of the $2^{128}$ possible solutions less than 2\\% is valid."]}
{"annotations": [{"id": 86736701, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}, {"id": 86901841, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86903623, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19125-66c30bd111f2a070.jsonl", "id_parag": "2310.19125-66c30bd111f2a070_4", "final_parag": "The SCRUM model of Fig.~\\ref{fig:scrumModel} is a software product line model defining the possible legal configurations of the SCRUM organizational framework. \nBILLING is a software product line model containing valid configurations for a billing software system. In both models, due to their constraints, less than 2\\% of its configurations are valid.", "candidate_comments": ["For example, the SCRUM model of Fig. \\ref{fig:scrumModel} which is an SPL defining the possible legal configurations of the SCRUM organization framework. As shown on  this figure, the SCRUM model contains 128 features and more than 250 constraints. These constraints are so complicated that out of the $2^{128}$ possible solutions less than 2\\% is valid.", "Another model from the SPLOT web site is BILLING. This is a simpler model that defines valid configurations for a billing software system. Containing 88 features and 166 constraints, BILLING is not as large as SCRUM. That said,   less than 2\\% of its $2^{88}$ configurations satisfy its internal constraints."]}
{"annotations": [{"id": 86733033, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86900169, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86900769, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19127-fe7b0e2d21a444e6.jsonl", "id_parag": "2310.19127-fe7b0e2d21a444e6_1", "final_parag": "{However, examining the representation of 235 PIEs that are largely unrelated in their literal and idiomatic context (their literal PIE embeddings and idiomatic definitions have a mean cosine similarity of 0.0047), we notice that their representations generated by the state-of-the-art \\cite{10.1162/tacl_a_00510} exhibit a high cosine similarity between their idiomatic and literal PIE embeddings (mean cosine similarity of 0.82).}", "candidate_comments": ["{However, examining the representation for 235 PIEs generated by the state-of-the-art \\cite{10.1162/tacl_a_00510}  reveals a high cosine similarity between the idiomatic and literal PIE embeddings (mean similarity of 0.82). In contrast, the mean cosine similarity between the literal PIE embeddings and their idiomatic definition embeddings is merely 0.0047, indicating that the PIEs' literal and idiomatic semantics are largely unrelated.} This means that the current SOTA has poor context-based semantic differentiation in the embedding space, despite generating good representations for idiomatic expressions. "]}
{"annotations": [{"id": 87251678, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87259370, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87267927, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19127-fe7b0e2d21a444e6.jsonl", "id_parag": "2310.19127-fe7b0e2d21a444e6_2", "final_parag": "Addressing these issues involves attending to the following challenges. (1) BART and GIEA's abilities should be combined to generate good embeddings for PIE in a context-dependent manner. (2) {With the self-supervised reconstruction task as the sole objective, PTLM parameters are already optimized for token reconstruction from their token embeddings. To represent PIEs, BART and GIEA's learning objectives should be revamped.}", "candidate_comments": ["Revamping the learning objective of both BART and GIEA, the embeddings of which are already optimized for token reconstruction from their token embeddings. With the self-supervised reconstruction task as the sole objective, the model has no motivation to alter its parameters.  (3) Unlike idiomatic PIEs, where their definitions are available, providing explicit embedding guidance to literal PIE is difficult. "]}
{"annotations": [{"id": 86981578, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86982944, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86983488, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19127-fe7b0e2d21a444e6.jsonl", "id_parag": "2310.19127-fe7b0e2d21a444e6_3", "final_parag": "To address these challenges, we propose \\textbf{P}otentially \\textbf{I}diomatic \\textbf{E}xpression \\textbf{R}epresentation generator (\\our). Inspired by AdapterFusion \\cite{pfeiffer-etal-2021-adapterfusion}, which has been used to combine task-specific adapters, \\our generates embeddings by combining the output from each GIEA adapter layer and pre-trained BART transformer layer with an attention fusion layer {serving as a routing mechanism that passes compositional or non-compositional embeddings based on the context.}.", "candidate_comments": ["These underlie the design of  \\textbf{P}otentially \\textbf{I}diomatic \\textbf{E}xpression \\textbf{R}epresentation generator (\\our) that takes as input a sentence with PIEs and automatically determines if the PIEs are literal (compositional) or idiomatic (non-compositional) based on the sentence context and generates semantically meaningful PIE embeddings."]}
{"annotations": [{"id": 86731398, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86731973, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86836370, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19127-fe7b0e2d21a444e6.jsonl", "id_parag": "2310.19127-fe7b0e2d21a444e6_4", "final_parag": "We implement an attention layer that acts on the outputs from each frozen GIEA's adapter layer and frozen BART's transformer layer and serves as a ``routing\" mechanism for the compositional and non-compositional type embeddings. To train the attention layer, {the overall loss is the sum of (1) a cosine similarity-based part for the embedding to encode meaning via external dictionary definitions, and (2) a reconstruction cross-entropy part that teaches the embedding of the association between the PIE senses and sentence contexts. Optimizing these two losses jointly allows the model to link PIEs' meanings to their contextual uses.}", "candidate_comments": ["\\SB{I think we should leave the objectives to below, but here we should talk at an abstract level of what we wish to achieve}   Specifically, we utilize a cosine similarity-based objective on top of reconstruction cross-entropy loss and we design a prompt infilling task to ask model to explicitly output the type (literal or idiomatic) or the definition of PIEs. "]}
{"annotations": [{"id": 86721656, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86725436, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86727055, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19127-fe7b0e2d21a444e6.jsonl", "id_parag": "2310.19127-fe7b0e2d21a444e6_5", "final_parag": "Specifically, given a sentence with a PIE at training time, we first generate the \\textit{PIE embedding} by mean pooling \\our final output embeddings of the PIE tokens. Then, we generate two embeddings that aid the {refining} of the PIE embedding: (1) we generate an \\textit{idiomatic embedding }that encodes the non-compositional meaning of the PIE by using MPNet \\cite{DBLP:conf/nips/Song0QLL20} to produce a sentence embedding on the PIE's idiomatic dictionary definition. We use MPNet because prior work \\cite{10.1162/tacl_a_00510} found the resulting definition embeddings help representation learning more than other models such as BART; and (2) we generate a \\textit{literal embedding} that encodes the compositional meaning of the PIE by mean pooling a regular PTLM's (here, BART) final PIE token embeddings. \n{Note that since the \"literal\" embeddings are contextualized, they may already encode idiomatic meanings for frequently used idioms, including idioms exclusively used figuratively. Our model should still provide accurate semantics for these idioms since the attention fusion layer could pass the idiomatic semantics through. However, for the vast majority of idioms that are rare in text, BART's embeddings are considered \"compositional,\" not capturing their figurative meanings, hence we refer to them as \\textit{literal} embeddings. \n}", "candidate_comments": ["\\ZZ{Note that since the \"literal\" embeddings are contextualized, it is possible that for the most frequent idioms in the pretraining corpus, the embeddings already encode their idiomatic meanings; the same goes for any idioms that are frequently used exclusively in their figurative sense (i.e., highly conventionalized). However, for these idioms, the output of our proposed model should remain semantically accurate. Moreover, for most idioms that are individually rare in text, we generally consider BART's embeddings to be \"compositional,\" as they do not capture the non-compositional, figurative meanings of idioms. Hence, we denote them as \\textit{literal} embeddings here.}"]}
{"annotations": [{"id": 86963290, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86985656, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86985683, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19127-fe7b0e2d21a444e6.jsonl", "id_parag": "2310.19127-fe7b0e2d21a444e6_6", "final_parag": "\\subsection{Prompt Infilling} To directly provide the PIE sense information to the model and help it relate PIE senses with sentence contexts, we design two types of prompt-based mask infilling tasks: (1) \\textit{type classification} prompts and (2) \\textit{definition generation} prompts.", "candidate_comments": ["Note that our use of prompts is unlike that , where prompts are used with trained LMs to perform specific tasks/%, such as classification, knowledge probing, and information extraction. "]}
{"annotations": [{"id": 86864903, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86899659, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86900183, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19127-fe7b0e2d21a444e6.jsonl", "id_parag": "2310.19127-fe7b0e2d21a444e6_7", "final_parag": "Note that unlike \\textit{prompt-based learning} \\cite{10.1145/3560815}, we use prompts to teach LMs informative representations.", "candidate_comments": ["Note that our use of prompts is unlike that , where prompts are used with trained LMs to perform specific tasks/%, such as classification, knowledge probing, and information extraction. "]}
{"annotations": [{"id": 86981912, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86983296, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86983997, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19127-fe7b0e2d21a444e6.jsonl", "id_parag": "2310.19127-fe7b0e2d21a444e6_8", "final_parag": "So, with its ability to produce meaningful idiomatic embeddings, GIEA achieves high sequence accuracy in SpanDET, whereas \\ourplus's ability to generate appropriate literal embeddings allows it to improve further.", "candidate_comments": [", which despite not improving recall, leads to a higher sequence accuracy that necessitates a precise literal and idiomatic sense disambiguation. "]}
{"annotations": [{"id": 86985105, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986330, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86986382, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19127-fe7b0e2d21a444e6.jsonl", "id_parag": "2310.19127-fe7b0e2d21a444e6_9", "final_parag": "\\textbf{Effect of Combined Components.} The salient effect of combining all components is shown by comparing FusionSim and \\our. \\our gains 16.5\\% in homogeneity score and 0.32 in inter-type cosine similarity while achieving 1.65\\% higher accuracy in SenseCLF and 17.3\\% higher sequence accuracy in SpanDET.", "candidate_comments": ["Adding the similarity learning objective, FusionSim gains 18.3\\% in homogeneity score and reduces 0.21 in inter-type cosine similarity. FusionSim's sequence accuracy for SpanDet also improved by 4.7\\% over that of FusionAttn. "]}
{"annotations": [{"id": 86980453, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture", "Réécriture", "Réécriture"]}, {"id": 86984143, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture", "Réécriture", "Réécriture"]}, {"id": 86984864, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture", "Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_1", "final_parag": "While image captioning models achieved notable benchmark performance in utilizing the correlation between visual and co-occurring labels to generate an accurate image description, this often results in a gender bias that relates to a specific gender, such as confidently identifying a woman when there is a kitchen in the image. The work of \\cite{zhao2017men} tackles the problem of gender bias in visual semantic role labeling by balancing the distribution. For image captioning, \\citet{hendricks2018women} consider ameliorating gender bias via a balance classifier. More recently, \\citet{hirota2022quantifying} measure gender bias amplification in image captioning via a trainable classifier on an additionally annotated existing caption dataset \\cite{zhao2021understanding} for the gender bias task, such as racial biases.", "candidate_comments": ["While image captioning models have shown promising benchmark results in utilizing the correlation between visual and co-occurring labels to generate an accurate image description, this often results in a gender bias that relates to a specific gender, such as confidently identifying a woman when there is a kitchen in the image. The work of \\cite{zhao2017men} tackles the problem of gender bias in visual semantic role labeling by balancing the distribution. For image captioning, \\citet{hendricks2018women} consider ameliorating gender bias via a balance classifier. More recently, \\citet{hirota2022quantifying} measure gender bias amplification in image captioning via a trainable classifier on an additionally annotated existing caption dataset \\cite{zhao2021understanding} for the gender bias task, such as racial biases.", "While image captioning models have shown promising results in utilizing the correlation between visual and co-occurring labels to generate an accurate static story of an image. This often results in a gender bias that relates to a specific gender, such as confidently identifying a woman when there is a kitchen in the image. The work of \\cite{zhao2017men} tackles the problem of gender bias in visual semantic role labeling by balancing the distribution. For image captioning, \\citet{hendricks2018women} consider ameliorating  gender bias via a balance classifier. More recently, \\citet{hirota2022quantifying} measure gender bias amplification in image captioning via a trainable classifier on an additionally annotated existing caption dataset \\cite{zhao2021understanding}  for the gender bias task, such as racial biases.", "Image captioning models utilize the correlation between visual and co-occurring labels to generate an accurate static story in an image. However, this often results in a gender bias that relates to a specific gender, such as confidently identifying a woman when there is a kitchen in the image. The work of \\cite{zhao2017men} tackles the problem of gender bias in visual semantic role labeling by constraining the model. For image captioning, \\citet{hendricks2018women} consider ameliorating  gender bias via a balance classifier. More recently, \\citet{hirota2022quantifying} measure gender bias amplification in image captioning via a trainable classifier on an additional annotated existing caption dataset \\cite{zhao2021understanding} for the gender bias task \\eg racial bias.", "original  Image captioning models utilize the correlation between visual and co-occurring labels to generate an accurate static story in an image. However, this often results in a gender bias that relates to a specific gender, such as confidently identifying a woman when there is a kitchen in the image. The work of \\cite{zhao2017men} tackles the problem of gender bias in visual semantic role labeling by balancing the distribution. For image captioning, \\citet{hendricks2018women} consider ameliorating  gender bias via a balance classifier. More recently, \\citet{hirota2022quantifying} measure gender bias amplification in image captioning via a trainable classifier on an additional annotated existing caption dataset \\cite{zhao2021understanding} for the gender bias task \\eg racial bias."]}
{"annotations": [{"id": 86870594, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86960965, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86958316, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_2", "final_parag": "To close the full picture of gender bias in image captioning, in this work, unlike other works, we examine the problem from a visual semantic relation perspective. We therefore propose a Gender Score via human-inspired judgment named Belief Revision \\cite{blok2003probability} which can be used to (1) discover bias and (2) predict gender bias without \\textit{training} or \\textit{unbalancing} the dataset. We conclude our contributions are as follows: (1) we investigate the gender object relation for image captioning at the word level (\\ie object-gender) and the sentence level with captions (\\ie gender-caption); (2) we propose a Gender Score that uses gender in relation to the visual information in the image to predict the gender bias. Our Gender Score can be used as a plug-in for any out-of-the-box image captioning system. Figure \\ref{fig:main_figure} shows an overview of the proposed gender bias measure for image captioning.", "candidate_comments": ["original To close the full picture of gender bias in image captioning, in this work, unlike other works, we examine the problem from a visual semantic relation perspective. We therefore propose a Gender Score via human inspired  judgment named Belief Revision \\cite{blok2003probability} which can be used to (1) discover bias and (2) predict gender bias without \\textit{unbalancing} the dataset. We conclude our contributions as follows: (1) We investigate the gender object relation for image captioning at the  word level (\\ie object-gender) and  the sentence level with captions (\\ie gender-caption); (2) we propose a gender score that uses gender in relation to the visual information in the image to predict the gender bias. Our Gender Score can be used as a plug-in for any out-of-the-box image captioning system. Figure \\ref{fig:main_figure} shows an overview of the proposed gender bias measure for caption generation."]}
{"annotations": [{"id": 86961084, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}, {"id": 86977372, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}, {"id": 86979017, "completed_by": {"id": 140156}, "result": ["Différent", "Différent"]}], "agreement": 55.55555555555556, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_3", "final_parag": "To obtain the \\textit{visual context} $o$ from each image $I$, we use out-of-the-box classifiers to extract the image context information $o(I)$. Specifically, following \\cite{sabir2023visual}, the objects extracted from all pre-trained models are obtained by extracting the top-3 object class/category (excluding \\textit{person} category) from each classifier after filtering out instances with (1) the cosine distance between objects (voting between classifiers of top-$k$) and (2) a low confidence score via a probability threshold $< 0.2$. We next describe each of these classifiers:", "candidate_comments": ["To obtain the \\textit{visual context} $o$ from each image $I$, we use out-of-the-box classifiers to extract the image context information $o(I)$. Specifically, following \\cite{sabir2023visual}, the objects extracted from all pre-trained models are obtained by extracting the top-3 object class/category  (excluding \\textit{person} category) from each classifier after filtering out instances with (1) the cosine distance between objects (voting between classifiers of top-$k$) and (2) a low confidence score via a probability threshold $< 0.2$. We next describe each of these classifier:", "To obtain the \\textit{visual context} $o$ from each image $I$, we use out-of-the-box classifiers to extract the image context information $o(I)$. Specifically, following \\cite{sabir2023visual}, the objects extracted from all pre-trained models are obtained by extracting the top-3 object class/category  (excluding \\textit{person} category) from each classifier after filtering out instances with (1) the cosine distance between objects (voting between classifiers of top-$k$) and (2) a low confidence score via a probability threshold $< 0.2$. We next describe each of these classifier:"]}
{"annotations": [{"id": 86960945, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86982884, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86983019, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent"]}], "agreement": 55.55555555555555, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_4", "final_parag": "{\\bf ResNet-152 \\cite{he2016deep}: }A residual deep network that is designed for ImageNet classification tasks, which relies heavily on batch normalization.", "candidate_comments": ["A residual deep network with many batch normalization layers that is designed for ImageNet classification tasks. \nThe shortcut is known as a gated recurrent unit that is able to ", "heavily on batch normalization A Residual Network (ResNet) is a deep neural network architecture designed for the ImageNet classification task that makes use of batch normalization."]}
{"annotations": [{"id": 86963056, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973422, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974672, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_5", "final_parag": "{\\bf Inception-ResNet FRCNN \\cite{huang2017speed}:} is an improved variant of Faster R-CNN, with a trade-off of better accuracy and fast inference via high-level features extracted from Inception-ResNet. It is a pre-trained model that is trained on COCO categories, with 80 object categories.", "candidate_comments": [", with 80 object categories. \n(excluding person label).  \"It is a pre-trained model that is trained on COCO categories, with 80 object categories.\""]}
{"annotations": [{"id": 86962484, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}, {"id": 86965112, "completed_by": {"id": 140156}, "result": ["Différent", "Réécriture"]}, {"id": 86970602, "completed_by": {"id": 126844}, "result": ["Différent", "Réécriture"]}], "agreement": 77.77777777777777, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_6", "final_parag": "The main component of our Gender Score (next section) is the semantic relation between the object and the gender. Therefore, in this section, we investigate the semantic correlation between the object and the gender in the general dataset (\\eg wiki). More specifically, we assume that all images contain a gender \\textit{man}, \\textit{woman} or gender neutral \\textit{person}, and we aim to measure the Cosine Distance between the object $o$, objects with context \\ie caption $y$ from the image $I$ and its related gender $a' \\in\\{\\text {man, woman, \\textit{person}}\\}$. In this context, we refer to the Cosine Distance as the closest distance (\\ie semantic relatedness score) between the gender-to-object via similarity measure \\eg cosine similarity. We apply our analysis to the most widely used human-annotated datasets in image captioning tasks: Flikr30K and COCO Captions datasets.", "candidate_comments": ["The main component of our Gender Score (next section) is the semantic relation between the object and the gender. Therefore, in this section, we investigate the semantic correlation between the object and the gender in the general dataset (\\eg wiki). More specifically,  we assume that all images contain a gender \\textit{man}, \\textit{woman} or gender neutral \\textit{person}, and we aim to measure the Cosine Distance between the object $o$, objects with context \\ie caption $y$ from the image $I$ and its related gender $a' \\in\\{\\text {man, woman, \\textit{person}}\\}$. In this context, we refer to the Cosine Distance as the closest distance (\\ie semantic relatedness score) between the gender-to-object via similarity measure  \\eg cosine similarity. We apply our analysis to the most widely used human-annotated datasets in image captioning tasks: Flikr30K and COCO Captions datasets.", "As shown in Table \\ref{tb:caption_result} original  The main component of our Gender Score (next section) is the semantic relation between the object and the gender. Therefore, in this section, we investigate the semantic correlation between the object and the gender in the general dataset (\\eg wiki). Specifically,  we assume that all images contain a gender \\textit{man}, \\textit{woman} or gender neutral \\textit{person}, and we aim to measure the Cosine Distance between the object $o$, object with context \\ie caption $y$ from the image $I$ and its related gender $a' \\in\\{\\text {man, woman, \\textit{person}}\\}$. In this context, we refer to the Cosine Distance as the closest distance (semantic relatedness score) between the gender-to-object via similarity measure  \\eg cosine similarity. We use the most widely employed human-annotated datasets in image captioning tasks: Flikr30K and COCO Captions. "]}
{"annotations": [{"id": 86961931, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86981652, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86981715, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_7", "final_parag": "We employ several commonly used pre-trained models: (1) word-level: GloVe \\cite{pennington2014glove}, fasttext \\cite{bojanowski2017enriching} and word2vec \\cite{mikolov2013efficient} out-of-the-box as baselines, then we utilize Gender Neutral GloVe \\cite{zhao2018learning}, which balances the gender bias; (2) sentence-level: Sentence-BERT \\cite{reimers2019sentence} tuned on Natural Language Inference (NLI) \\cite{conneau2017supervised}, (2b) SimCSE \\cite{gao2021simcse} contrastive learning supervised by NLI, and an improved SimCSE version with additional \\textbf{Info}rmation aggregation via masked language model objective InfoCSE \\cite{wu2022infocse}. In particular, we measure the Cosine Distance from the gender $a'$ to the object $o$ or caption $y$ as shown in Figure \\ref{fig:main_figure} (a, b) (\\eg how close the gender vector of $a'$ is to the object $o$ \\textit{umbrella}). Table \\ref{tb:glove} shows the top most object-gender frequent count with the standard and gender-balanced GloVe, respectively.", "candidate_comments": ["without RoBERTa original  We employ a number of common pre-trained models: (1) word-level: GloVe \\cite{pennington2014glove}, fasttext \\cite{bojanowski2017enriching} and word2vec \\cite{mikolov2013efficient} out-of-the-box as baselines, then we employ Gender Neutral GloVe \\cite{zhao2018learning}, which balances the gender bias; (2) Sentence-level: Sentence-BERT \\cite{reimers2019sentence} tuned on Natural Language Inference  (NLI) \\cite{conneau2017supervised}, (2b) SimCSE \\cite{gao2021simcse} contrastive learning supervised by NLI, and an improved SimCSE version with additional \\textbf{Info}rmation aggregation via masked language model objective InfoCSE \\cite{wu2022infocse}. In particular, we measure the Cosine Distance from the gender $a'$ to the object $o$ or caption $y$ as shown in Figure \\ref{fig:main_figure} (a, b) (\\eg how close the gender vector of $a'$ is to the object $o$, \\textit{umbrella}). Table \\ref{tb:glove} shows the top most frequent count object-gender with standard and gender balance GloVe, respectively."]}
{"annotations": [{"id": 86972455, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86983822, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}, {"id": 86988686, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_8", "final_parag": "where $m$ and $w$ refer to the gender in the image, and the $s$ is our proposed gender-to-object semantic relation bias score. In our case, we also use the score to compute the ratio to gender neutral \\textit{person}:", "candidate_comments": ["original where \\textbf{m}en and \\textbf{w}omen refer to the visual in the image, \\textbf{s}core is our proposed gender-to-object semantic relation bias score, and the \\textbf{c}ount is the co-occurrence with the \\textbf{obj}ect as a pair in the dataset. In our case, we also use the score to compute the ratio to \\textit{person}: ", "original  where \\textbf{m}en and \\textbf{w}omen refer to the gender in the image, \\textbf{s}core is our proposed gender-to-object semantic relation bias score, and the \\textbf{c}ount is the co-occurrence with the context \\textbf{obj}ect as a pair in the dataset. In our case, we also use the score to compute the ratio to gender neutral  \\textit{person}: "]}
{"annotations": [{"id": 86722208, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture", "Réécriture", "Réécriture"]}, {"id": 86726880, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture", "Réécriture", "Réécriture"]}, {"id": 86725127, "completed_by": {"id": 62306}, "result": ["Réécriture", "Différent", "Réécriture", "Réécriture"]}], "agreement": 86.66666666666666, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_9", "final_parag": "In this section, we describe the proposed \\textbf{G}ender \\textbf{S}core that estimates gender based on its semantic relation with the visual information extracted from the image. \\citet{sabir2022belief} proposed a caption re-ranking method that leverages visual semantic context. This approach utilized Belief Revision \\cite{blok2003probability} to convert the similarity \\ie Cosine Distance (Section \\ref{sec3:cosine}) into a probability measure.", "candidate_comments": ["original  In this section, we describe the proposed \\textbf{G}ender \\textbf{S}core that estimates gender based on its semantic relation  with the visual information extracted from the image. \\citet{sabir2022belief} proposed a caption re-ranker with visual semantic context via similarity to probability. This approach utilized Belief Revision \\cite{blok2003probability} to convert the similarity (\\ie Cosine Distance) to probability. % with human judgment.  ", "CR long  In this section, we describe the proposed \\textbf{G}ender \\textbf{S}core that estimates gender based on its semantic relation  with the visual information extracted from the image. \\citet{sabir2022belief} proposed a caption re-ranking method that leverages visual semantic context.  This approach utilized Belief Revision \\cite{blok2003probability} to convert the similarity (\\ie Cosine Distance ) into a probability measure.", " CR long  In this section, we describe the proposed \\textbf{G}ender \\textbf{S}core that estimates gender based on its semantic relation  with the visual information extracted from the image. \\citet{sabir2022belief} proposed a caption re-ranker via visual semantic context. This approach utilized Belief Revision \\cite{blok2003probability} to convert the similarity between the caption and the visual context (\\ie Cosine Distance) to probability.", "n this section, we introduce the concept of the Gender Score (GS), which estimates gender based on its semantic connection with visual information extracted from images. Sabir et al. (2022) presented a caption re-ranking method that leverages visual semantic context. This innovative approach applied Belief Revision (Blok et al., 2003) to transform the similarity between the caption and the visual context (i.e., Cosine Distance) into a probability measure."]}
{"annotations": [{"id": 86730625, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture", "Différent"]}, {"id": 86731342, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture", "Différent"]}, {"id": 86766136, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture", "Différent"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_10", "final_parag": "{\\textbf{Belief Revision.}} To obtain likelihood bias revisions based on similarity scores (\\ie gender, object), we need three parameters: (1) \\textbf{Hypothesis (g)}: caption $y$ with the associated gender $a \\in\\{\\text {man, woman}\\}$, (2) \\textbf{Informativeness (c)}: image object information $ o$ confidence and (3) \\textbf{Similarities}: the degree of relatedness between object and gender $sim(y,o)$.", "candidate_comments": ["{\\textbf{Belief Revision}} To obtain the likelihood revisions based on similarity scores (\\ie gender, object), we need three parameters: (1) \\textbf{Hypothesis}: prior probabilities, (2) \\textbf{Informativeness}: conclusion events and (3) \\textbf{Similarities}: the degree of relatedness between involved categories as ", "{\\textbf{Belief Revision.}} To obtain the likelihood revisions based on similarity scores (\\ie gender, object), we need three parameters: (1) \\textbf{Hypothesis}: caption with the associated gender, (2) \\textbf{Informativeness}: image object information and (3) \\textbf{Similarities}: the degree of relatedness between  object and gender. The object-gender bias revision for Gender Score:  ", "original  {\\textbf{Belief Revision.}} To obtain likelihood bias revisions based on similarity scores (\\ie gender, object), we need three parameters: (1) \\textbf{Hypothesis (g)}: caption $y$ with the associated gender $a \\in\\{\\text {man, woman}\\}$, (2) \\textbf{Informativeness (c)}: image object information $ o$ confidence and (3) \\textbf{Similarities}: the degree of relatedness between  object and gender $sim(y,o)$.  "]}
{"annotations": [{"id": 86726343, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86729306, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}, {"id": 86825866, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_11", "final_parag": "where $\\operatorname{P}(g_{y})$ is the \\textit{hypothesis} probability of $y$, $\\mathcal{D}$ is the predicted captions with the gender $a$, and $\\operatorname{P}(c_{o})$ is the probability of the evidence that causes hypothesis probability revision \\ie visual bias", "candidate_comments": ["where $\\operatorname{P}(w)$ is the \\textit{hypothesis} probability (word or caption) and  $\\operatorname{P}(c)$ is the probability of the evidence that causes hypothesis probability revision (visual context  from the image).  We discuss the detail of each component next,", "original short where $\\operatorname{P}(g)$ is the \\textit{hypothesis} probability (caption with the gender) and  $\\operatorname{P}(c)$ is the probability of the evidence that causes hypothesis probability revision (\\ie context object from the image) as: "]}
{"annotations": [{"id": 86951787, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86958327, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86959911, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_12", "final_parag": "{\\textbf{Hypothesis}:} $\\operatorname{P}(g_{y}) \\quad (\\small{\\text{caption}} \\, y \\, \\small{\\text{with the gender}}\\, a)$", "candidate_comments": ["{\\textbf{Hypothesis}:} $\\operatorname{P}(g)\\rightarrow$   {\\small{caption $y$ with the gender $a$}} %\\texttt{\\small candidate caption} "]}
{"annotations": [{"id": 86964726, "completed_by": {"id": 140156}, "result": ["Différent", "Différent"]}, {"id": 86972644, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}, {"id": 86973751, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_13", "final_parag": "{\\textbf{Informativeness}:} $1-\\operatorname{P}(c_{o}) \\quad (\\small{\\text{object context}} \\, o$)", "candidate_comments": ["{\\textbf{Hypothesis}:} $\\operatorname{P}(g)\\rightarrow$   {\\small{caption $y$ with the gender $a$}} %\\texttt{\\small candidate caption} ", "{\\textbf{Similarities}:}  $\\alpha=\\left[\\frac{1-\\operatorname{sim}(y, o)}{1+\\operatorname{sim}(y, o)}\\right]^{1-\\operatorname{P}(c)} (o(I)\\, \\small{\\text{{bias wight}}})$"]}
{"annotations": [{"id": 86962607, "completed_by": {"id": 62471}, "result": ["Différent", "Différent", "Différent", "Différent"]}, {"id": 86979489, "completed_by": {"id": 126844}, "result": ["Différent", "Différent", "Différent", "Différent"]}, {"id": 86979865, "completed_by": {"id": 140156}, "result": ["Différent", "Différent", "Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_14", "final_parag": "{\\textbf{Similarities}:} $\\alpha=\\left[\\frac{1-\\operatorname{sim}(y, o)}{1+\\operatorname{sim}(y, o)}\\right]^{1-\\operatorname{P}(c_{o})} (\\small{\\text{{visual bias}}})$", "candidate_comments": ["\\quad (\\text{Newton's second law}) {\\textbf{Similarities}:}  $\\alpha=\\left[\\frac{1-\\operatorname{sim}(g, c)}{1+\\operatorname{sim}(g, c)}\\right]^{1-\\operatorname{P}(c)}$  %\\texttt{\\small (visual relatedness to the gender)}", "{\\textbf{Similarities}:}  $\\alpha=\\left[\\frac{1-\\operatorname{sim}(g, c)}{1+\\operatorname{sim}(g, c)}\\right]^{1-\\operatorname{P}(c)}$  %\\texttt{\\small (visual relatedness to the gender)} {\\textbf{Similarities}:}  $\\alpha=\\left[\\frac{1-\\operatorname{sim}(y, o)}{1+\\operatorname{sim}(y, o)}\\right]^{1-\\operatorname{P}(c)}\\rightarrow$ {\\small{obj   weight}}", "{\\textbf{Similarities}:}  $\\alpha=\\left[\\frac{1-\\operatorname{sim}(y, o)}{1+\\operatorname{sim}(y, o)}\\right]^{1-\\operatorname{P}(c)} (o(I)\\, \\small{\\text{{bias wight}}})$", "{\\textbf{Similarities}:}  $\\alpha=\\left[\\frac{1-\\operatorname{sim}(y, o)}{1+\\operatorname{sim}(y, o)}\\right]^{1-\\operatorname{P}(c)} (o(I)\\, \\small{\\text{{bias}}})$ {\\textbf{Similarities}:}  $\\alpha=\\left[\\frac{1-\\operatorname{sim}(y, o)}{1+\\operatorname{sim}(y, o)}\\right]^{1-\\operatorname{P}(c)}$ {(\\small{object   weight})}"]}
{"annotations": [{"id": 86899023, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}, {"id": 86902191, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86906590, "completed_by": {"id": 140156}, "result": ["Différent", "Réécriture"]}], "agreement": 77.77777777777777, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_15", "final_parag": "{The visual} context $\\operatorname{P}(c_{o})$ will revise the caption with the associated gender $\\operatorname{P}(g_{y})$ (\\ie gender bias) if there is a semantic relation between them $sim(y,o)$. We discuss each component next:", "candidate_comments": ["{The visual} context $c$ will revise the caption with the associated gender $g$ (\\ie gender bias) if there is semantic relation between them $sim(g,c)$. We discuss the detail of each component next: ", "original  {The visual} context $c$ will revise the caption with the associated gender $g$ (\\ie gender bias) if there is semantic relation between them $sim(y,o)$. We discuss the detail of each component next: "]}
{"annotations": [{"id": 86730991, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture", "Réécriture"]}, {"id": 86726027, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture", "Réécriture"]}, {"id": 86729859, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_16", "final_parag": "{\\textbf{Hypothesis initial bias}:} In visual-based belief revision, one of the conditions is to start with an initial hypothesis and then revise it using visual context and a similarity score. Therefore, we initialize the caption hypothesis $\\operatorname{P}(g_{y})$ with a common observation $\\operatorname{P}(g'_{y})$, such as a Language Model (LM) (we consider this as an initial bias without visual). We employ (GPT-2) \\cite{radford2019language} with mean token probability since it achieves better results.", "candidate_comments": ["fixed {\\textbf{Hypothesis initial bias}:} As one of the belief revision conditions, we need an initial hypothesis and then apply the visual context to revise the gender relatedness score. Therefore, we initialize the caption hypothesis  $\\operatorname{P}(g_{y})$ with a common observation $\\operatorname{P}(g'_{y})$, such as a Language Model (LM) (we consider this as an initial bias without visual). We employ (GPT-2) \\cite{radford2019language} with mean token probability since it achieves better results. ", "{\\textbf{Hypothesis initial bias}:} In visual belief revision, one of the conditions is to start with an initial hypothesis and then apply the visual context to revise it via  similarity score. Therefore, we initialize the caption hypothesis  $\\operatorname{P}(g_{y})$ with a common observation $\\operatorname{P}(g'_{y})$, such as a Language Model (LM) (we consider this as an initial bias without visual). We employ (GPT-2) \\cite{radford2019language} with mean token probability since it achieves better results. ", "{\\textbf{Hypothesis initial bias}:} As the belief revision is a human-inspired model, we need an initial hypothesis and then apply the visual context to revise the gender relatedness score. Therefore, we initialize the caption hypothesis  $\\operatorname{P}(g_{y})$ with a common observation $\\operatorname{P}(g'_{y})$, such as a Language Model (LM) (we consider this as an initial bias without visual). We employ (GPT-2) \\cite{radford2019language} with mean token probability since it achieves better results. "]}
{"annotations": [{"id": 86736119, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture", "Réécriture", "Réécriture"]}, {"id": 86837908, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture", "Réécriture", "Réécriture"]}, {"id": 86898234, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture", "Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_17", "final_parag": "{\\bf Informativeness of bias information:} As the visual context probability $\\operatorname{P}(c_{o})$ approaches 1 and in consequence is less informative (very frequent objects have no discriminative power since they may co-occur with any gender) $1-\\operatorname{P}(c_{o})$ approaches zero, causing $\\alpha$ to get closer to 1, and thus, a smaller revision of $\\operatorname{P}(g'_{y})$. Therefore, as we described in the Visual Context Information Section \\ref{se:data}, we leverage a threshold and semantic filter visual context dataset from ResNet, Inception-ResNet v2 based Faster R-CNN object detector, and CLIP to extract top-$k$ textual visual context information from the image. The extracted object is used to measure the gender-to-object bias direct relation.", "candidate_comments": ["{\\bf Informativeness:} As we described in the Visual Context Information Section \\ref{se:data}, we leverage ResNet, Inception-ResNet v2 based Faster R-CNN object detector \\footnote{TensorFlow Object Detection API} and CLIP to extract Top-$k$ three textual visual context information from the image. We take the mean probability of the three visual context.", "{\\bf Informativeness of bias information:} As the visual context probability $\\operatorname{P}(c)$ approaches 1 and in consequence is less informative (very frequent objects have no discriminative power since they may co-occur with any gender) $1-\\operatorname{P}(c)$  approaches zero, causing $\\alpha$ to get closer to 1, and thus, a smaller revision of $\\operatorname{P}(g')$. Therefore, as we described in the Visual Context Information Section \\ref{se:data}, we leverage a threshold and semantic filter visual context dataset from  ResNet, Inception-ResNet v2 based Faster R-CNN object detector, and CLIP to extract Top-$k$ three textual visual context information from the image. The extracted object is used to measure the gender-to-object bias direct  relation. ", "As the probability of the visual context P(c) approaches 1 and in consequence is less informative (very frequent objects have no discriminative power since they may co-occurr with anything) 1 − P (c) approaches zero, causing α to get closer to 1, and thus, a smaller revision of P (w).", "As the probability of the visual context P(c) approaches 1 and in consequence is less informative (very frequent objects have no discriminative power since they may co-occurr with anything) 1 − P (c) approaches zero, causing α to get closer to 1, and thus, a smaller revision of P (w)."]}
{"annotations": [{"id": 86904224, "completed_by": {"id": 62471}, "result": ["Différent", "Réécriture", "Réécriture"]}, {"id": 86972738, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture", "Réécriture"]}, {"id": 86974838, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture", "Réécriture"]}], "agreement": 83.33333333333334, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_18", "final_parag": "{\\bf Relatedness between hypothesis and bias information:} Likelihood revision occurs if there is a close correlation between the hypothesis and the new information. As the $sim(y,o)$ (gender, object), gets closer to 1 (higher relatedness) $\\alpha$ gets closer to 0, and thus hypothesis probability is revised (\\ie gender bias) and raised closer to 1. Therefore, the initial hypothesis will be revised or backed off to 1 (no bias) based on the relatedness score. In our case, we employ Sentence-BERT to compute the Cosine Distance (Section \\ref{sec3:cosine}) by using object $o$ as context for the caption $y$ with associated gender $a$.", "candidate_comments": ["{\\bf Similarities:} Hypothesis revision is more likely if there is a close relation between the hypothesis and the new information (candidate caption and visual context in our case). We rely on two of the most recent state-of-the-art pre-trained Transformer-based language models to compute the semantic similarity between the caption and its visual context information with contextual embedding:", "{\\bf Similarities:} Likelihood revision occurs if there is a close correlation between the hypothesis and the new information. As the $sim(w,c)$ (object, gender), gets closer to 1 (higher relatedness) $\\alpha$ get closer to 0 and thus hypothesis probability is revised (\\ie gender bias) and raised closer to 1. In our case, we employ Sentence-BERT to compute the Cosine Distance (Sec \\ref{sec3:cosine}) by using the objects as the context for the sentence with gender. ", "{\\bf Similarities:} Likelihood revision occurs if there is a close correlation between the hypothesis and the new information. As the $sim(w,c)$ (object, gender), gets closer to 1 (higher relatedness) $\\alpha$ get closer to 0 and thus hypothesis probability is revised (\\ie gender bias) and raised closer to 1. In our case, we employ Sentence-BERT to compute the Cosine Distance (Sec \\ref{sec3:cosine}) by using the objects as the context for the sentence with gender. "]}
{"annotations": [{"id": 86736542, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86959496, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86960467, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_19", "final_parag": "{\\bf Caption model.} We examine seven of the most recent Transformer state-of-the-art caption models: Transformer \\cite{Ashish:17} with bottom-up top-down features \\cite{anderson2018bottom}, AoANet \\cite{huang2019attention}, Vilbert \\cite{lu202012}, OSCAR \\cite{li2020oscar}, BLIP \\cite{li2022blip}, Transformer with Reinforcement Learning (RL) CLIPS+CIDEr as image+text similarity Reward (TraCLIPS-Reward) \\cite{cho2022fine} and Large LM$_{\\text{2.7B}}$ based BLIP-2 \\cite{li2023blip}. Note that for a fair comparison with other pre-trained models, OSCAR uses a cross-entropy evaluation score rather than the RL-based CIDEr optimization score.", "candidate_comments": ["original {\\bf Caption model.} We examine seven of the most recent Transformer  state-of-the-art caption models: Transformer \\cite{Ashish:17} with up bottom up features \\cite{anderson2018bottom}, AoANet \\cite{huang2019attention} , Vilbert  \\cite{lu202012}, OSCAR  \\cite{li2020oscar}, BLIP \\cite{li2022blip}, Transformer with Reinforcement Learning (RL) CLIPS+CIDEr as image+text similarity Reward (TraCLIPS-Reward) \\cite{cho2022fine} and Large LM$_{\\text{2.7B}}$ based BLIP-2 \\cite{li2023blip}. Note that, for fair comparison all pre-trained models are using cross-entropy evaluation score rather than the RL-based CIDEr optimization score. "]}
{"annotations": [{"id": 86963160, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}, {"id": 86972016, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}, {"id": 86972153, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_20", "final_parag": "Our experiments apply visual semantics between the object or object with context \\ie caption and its related gender information to predict object-to-gender related bias. The proposed visual-gender scores are (1): Cosine Distance, which uses the similarity score to try to estimate the proper gender as shown in Table \\ref{tb:caption_result}; (2) Gender Score, which carries out Belief Revision (visual bias likelihood revision) to revise the hypothesis initial bias via similarity \\eg Cosine Distance (gender, object). For our baseline, we adopt the Object Gender Co-Occ metric \\cite{zhao2017men} for the image captioning task.", "candidate_comments": ["Our experiments apply visual semantics between the visual or caption and its related gender information to predict the related gender. The proposed visual-gender score are (1): Cosine Distance that uses the similarity score to try to estimate the proper gender as shown in  Table~\\ref{tb:caption_result}. (2) Gender Score that carries out the Belief Revision (likelihood revision) to revise the hypothesis via similarity (Gender, visual). ", "original  Our experiments apply visual semantics between the visual or caption and its related gender information to predict the related gender. The proposed visual-gender scores are (1): Cosine Distance, which uses the similarity score to try to estimate the proper gender as shown in Table \\ref{tb:caption_result}; (2) the Gender Score, which carries out the Belief Revision (likelihood revision) to revise the hypothesis via similarity \\eg Cosine Distance (gender, object). For our baseline, we adopt Object Gender Co-Occ \\cite{zhao2017men} for the image captioning task.  "]}
{"annotations": [{"id": 86971125, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture", "Réécriture"]}, {"id": 86973276, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture", "Réécriture"]}, {"id": 86973845, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_21", "final_parag": "In this work, we hypothesize that every image has a gender $\\in\\{\\text {man, woman}\\}$ or gender neutral (\\ie \\textit{person}), and our rationale is that each model suffers from \\textit{Right for the Wrong Reasons} \\cite{hendricks2018women} or leakage \\cite{wang2019balanced} (see Table \\ref{tb:caption_test}), such as associating all kitchens with women. Therefore, we want to explore all the cases and let the proposed distance/score decide which gender (\\ie bias) is in the image based on a visual bias. In particular, inspired by the cloze probability last word completion task \\cite{gonzalez2007methods}, we generate two identical sentences but with a different gender, and then we compute the likelihood revisions between the sentence-gender and the caption using the object probability. Table~\\ref{tb:caption_result} shows that GloVe and GN-GloVe (balanced) have identical results on COCO Captions dataset, which indicate that not all objects have a strong bias toward a specific gender. In addition, Table~\\ref{tb:caption_test} shows that our score balances the bias of the Cosine Distance and demonstrates, on average, that not all objects have a strong gender bias. Also, our approach detects strong specific object-to-gender bias and has a similar result to the existing Object Gender Co-Occ method on the most biased object toward men, as shown in Table~\\ref{tb:caption-example}. TraCLIPS-Reward inherits biases from RL-CLIPS and thus generates caption w/o a specific gender (\\eg \\textit{person}, \\textit{guy}, etc). Therefore, we adopt the combined CLIPS+CIDEr Rewards, which suffer less gender prediction error.", "candidate_comments": ["new In this work, we hypothesize that every image has a gender (\\ie \\textit{man} or \\textit{woman}) or neutral gender (\\ie \\textit{person}), and our rationale is that each model suffers from \\textit{Right for the Wrong Reasons} or leakage \\cite{wang2019balanced} (see Table \\ref{tb:caption_test}), such as associating all kitchen with women. Therefore, we want to explore all the cases and let the proposed distance/score decide which gender (\\ie bias) is in the image based on a visual bias. In particular, inspired by the cloze probability last word completion task \\cite{gonzalez2007methods}, we generate two identical sentences but with different gender, and then we compute the likelihood revisions between the sentence-gender and the caption using the object probability. Table~\\ref{tb:caption_test} shows that our score balances the bias of the Cosine Distance, and demonstrates, on average, not all objects have a strong gender bias. Also, our approach detects strong specific object-to-gender bias and has a similar result to the existing Object Gender Co-Occ method \\cite{zhao2017men} on the most biased object toward men, as shown in Table~\\ref{tb:caption-example}.", "original  In this work, we hypothesize that every image has a gender $\\in\\{\\text {man, woman}\\}$ or gender neutral  (\\ie \\textit{person}), and our rationale is that each model suffers from \\textit{Right for the Wrong Reasons} \\cite{hendricks2018women} or Leakage \\cite{wang2019balanced} (see Table \\ref{tb:caption_test}), such as associating all kitchens with women. Therefore, we want to explore all the cases and let the proposed distance/score decide which gender (\\ie bias) is in the image based on a visual bias. In particular, inspired by the cloze probability last word completion task \\cite{gonzalez2007methods}, we generate two identical sentences but with a different gender, and then we compute the likelihood revisions between the sentence-gender and the caption using the object probability. Table~\\ref{tb:caption_test} shows that our score balances the bias of the Cosine Distance and demonstrates, on average, that not all objects have a strong gender bias. Also, our approach detects strong specific object-to-gender bias and has a similar result to the existing Object Gender Co-Occ method on the most biased object toward men, as shown in Table~\\ref{tb:caption-example}. TraCLIPS-Reward (best proposed model aka CLIPS+Grammar reward) inherits biases from RL and thus generates caption w/o a specific gender (\\eg  \\textit{person}, \\textit{guy}, etc). Therefore, we adopt the combined CLIPS+CIDEr Rewards, which suffer less gender prediction error.   ", "In this work, we hypothesize that every image has a gender (\\ie \\textit{man} or \\textit{woman}) or neutral gender (\\ie \\textit{person}), and our rationale is that each model suffers from \\textit{Right for the Wrong Reasons} or leakage \\cite{wang2019balanced}, such as associating all food with a woman. In particular, inspired from  cloze probability  last word complication task \\cite{gonzalez2007methods} we generate two identical sentences but with different gender, and then we compute the semantic relation between the sentence-gender and the caption. Table~\\ref{tb:caption_test} shows that our gender score balances the bias of the cosine distance, and demonstrates that in average not all objects have a strong gender bias. Also, our approach detects strong specific object-to-gender bias and has a similar result to exiting Object Gender Co-occurring method on the most biased object toward men, as shown in Table~\\ref{tb:caption-example}."]}
{"annotations": [{"id": 86960406, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture", "Réécriture", "Réécriture", "Réécriture"]}, {"id": 86971847, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture", "Réécriture", "Réécriture", "Réécriture"]}, {"id": 86972080, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture", "Réécriture", "Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_22", "final_parag": "{\\bf{Gender Score Estimation.}} In this experiment, we < \\texttt{MASK} > the gender and use the object with context to predict the gender as shown in Figure \\ref{tb-fig:gender-est}. The idea is to measure the amplified gender-to-object bias in pre-trained models. Table \\ref{tb:benh-mark} shows that the fill-in gender score has a more bias towards men results than object-gender pair counting. The rationale is that the model can estimate the strong gender object bias as shown in Figure \\ref{tb-fig:gender-est}, including the false positive and leakage cases by the classifier.", "candidate_comments": ["original  {\\bf{Gender Score Estimation.}} In this experiment, we \\texttt{Mask} the gender and use the object to predict the gender as shown in Figure \\ref{tb-fig:gender-est}. The idea is to measure the bias-relatedness to the object.  Table  \\ref{tb:benh-mark} shows that the gender score has a more accurate bias towards man results than Object gender pair counting. Especially when the gender-to-object bias is: 1) is not strong as shown in Figure \\ref{tb-fig:gender-est} with \\textit{tennis},  2) misrecognized by the classifier.% (false positive or negative). %Also our appraoch  balance the bias when  as shown in Figure \\ref{}", "{\\bf{Gender Score Estimation.}} In this experiment, we \\texttt{Mask} the gender and use the object to predict the gender as shown in Figure \\ref{tb-fig:gender-est}. The idea is to measure the bias-relatedness to the object.  Table  \\ref{tb:benh-mark} shows that the gender score has a more accurate bias towards man results than, our baseline, object gender pair counting. Especially when the gender-to-object bias: 1) is not strong as shown in Figure \\ref{tb-fig:gender-est} with \\textit{tennis},  2) misrecognized or when there is leakage by the classifier.", "{\\bf{Gender Score Estimation.}} In this experiment, we \\texttt{Mask} the gender and use the object to predict the gender as shown in Figure \\ref{tb-fig:gender-est}. The idea is to measure the gender bias-relatedness to the object.  Table  \\ref{tb:benh-mark} shows that the predictive gender score has a more accurate  bias towards man results than, our baseline, object gender pair counting. Especially when the gender-to-object bias: 1) is not strong as shown in Figure \\ref{tb-fig:gender-est} with \\textit{tennis},  2) misrecognized or when there is leakage by the classifier.", "{\\bf{Gender Score Estimation.}} In this experiment, we \\texttt{Mask} the gender and use the object to predict the gender as shown in Figure \\ref{tb-fig:gender-est}. The idea is to measure the amplified gender-to-object bias in pre-trained models. Table  \\ref{tb:benh-mark} shows that the fill-in gender score has a more  bias towards man results than human annotations  and object gender pair counting. Especially when the gender-to-object bias: 1) is not strong as shown in Figure \\ref{tb-fig:gender-est} with \\textit{tennis},  2) misrecognized or when there is leakage by the classifier.", "original long version  {\\bf{Gender Score Estimation.}} In this experiment, we <\\texttt{Mask}> the gender and use the object to predict the gender as shown in Figure \\ref{tb-fig:gender-est}. The idea is to measure the amplified gender-to-object bias in pre-trained models. Table  \\ref{tb:benh-mark} shows that the fill-in gender score has a more  bias towards man results than human annotations  and object gender pair counting. The reason is that our our approach measures the amplified gender bias more accurately than the Object Gender Co-Occ in the following two scenarios: (1) where the gender is not obvious in the image and is misclassified by the caption baseline, and (2) when there is leakage by the classifier. In addition, unlike the Object Gender Co-Occ, our model balances the gender to object bias and only measures a strong object to gender bias relation as shown in Figure \\ref{tb-fig:gender-est}. Therefore, our approach can be used in the aforementioned scenarios as a stand-alone metric or as an additional metric alongside the existing Object Gender Co-Occ approach."]}
{"annotations": [{"id": 87251737, "completed_by": {"id": 70661}, "result": ["Différent", "Réécriture"]}, {"id": 87259279, "completed_by": {"id": 126844}, "result": ["Différent", "Réécriture"]}, {"id": 87268359, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}], "agreement": 77.77777777777777, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_23", "final_parag": "{\\bf{Discussion.}} Our approach measures the amplified gender bias more accurately than the Object Gender Co-Occ \\cite{zhao2017men} in the following two scenarios: (1) where the gender is not obvious in the image and is misclassified by the caption baseline, and (2) when there is leakage by the classifier. In addition, unlike the Object Gender Co-Occ, our model balances the gender to object bias and only measures a strong object to gender bias relation as shown in Figure \\ref{tb-fig:gender-est}. For instance, the word “hitting” in a generated caption (as a stand-alone without context) is associated with a bias toward men more than women, and it will influence the final gender-to-caption bias score. However, our gender score balances the unwanted bias and only measures the pronounced gender to object bias relation.", "candidate_comments": ["original long  {\\bf{Discussion.}} Our approach measures the amplified gender bias more accurately than the Object Gender Co-Occ \\cite{zhao2017men} in the following two scenarios: (1) where the gender is not obvious in the image and is misclassified by the caption baseline, and (2) when there is leakage by the classifier. In addition, unlike the Object Gender Co-Occ, our model balances the gender to object bias and only measures a strong object to gender bias relation as shown in Figure \\ref{tb-fig:gender-est}. For instance, the word “hitting” in a generated caption (as a stand-alone without context) is associated with a bias toward men more than women, and it will influence the final gender-to-caption bias score. However, our gender score balances the unwanted bias and only measures the pronounced gender to object bias relation. Therefore, our gender score can be used in the aforementioned scenarios as a stand-alone metric or as an additional metric alongside the existing Object Gender Co-Occ approach. ", "original long version  {\\bf{Gender Score Estimation.}} In this experiment, we <\\texttt{Mask}> the gender and use the object to predict the gender as shown in Figure \\ref{tb-fig:gender-est}. The idea is to measure the amplified gender-to-object bias in pre-trained models. Table  \\ref{tb:benh-mark} shows that the fill-in gender score has a more  bias towards man results than human annotations  and object gender pair counting. The reason is that our our approach measures the amplified gender bias more accurately than the Object Gender Co-Occ in the following two scenarios: (1) where the gender is not obvious in the image and is misclassified by the caption baseline, and (2) when there is leakage by the classifier. In addition, unlike the Object Gender Co-Occ, our model balances the gender to object bias and only measures a strong object to gender bias relation as shown in Figure \\ref{tb-fig:gender-est}. Therefore, our approach can be used in the aforementioned scenarios as a stand-alone metric or as an additional metric alongside the existing Object Gender Co-Occ approach."]}
{"annotations": [{"id": 86982956, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86986717, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86986800, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_24", "final_parag": "The accuracy of our Gender Score heavily relies on the semantic relation (\\ie Cosine Distance) between gender and object, a high or low degree of similarity score between them can influence the final bias score negatively. Also, our model relies on a large pre-trained model(s), which inherently encapsulate their own latent biases that might impact the visual bias revision behavior in gender-to-object bias scenarios. Specifically, in cases when there are multiple strong bias contexts (\\ie non-objects context) with high similarity scores toward a particular gender present within the caption. This can imbalance the final gender-to-object score, leading to errors in gender prediction and bias estimation. In addition, the false positive object context information extracted by the visual classifier will result in inaccurate bias estimation.", "candidate_comments": ["As the gender score relies heavily on the semantic relation (\\ie Cosine Distance) between gender and object, a  high or low degree of similarity can unbalance the score.  Also, the false positive object information by the visual classifier influences the score negatively which lead to wrong bias estimation.  As the gender score relies heavily on the semantic relation (\\ie Cosine Distance) between gender and object, a  high or low degree of similarity can unbalance the score.  Also, our model relies on a large pre-trained model(s), which carry their own inherent biases that might influence the bias score. For instance, if there is multiple bias context toward a particular gender in the caption, the final gender-to-object score will be diminished, which results in gender prediction and bias error. Another limitation is the false positive object information by the visual classifier influences the score negatively, which results in wrong bias estimation.", "As the gender score relies heavily on the semantic relation (\\ie Cosine Distance) between gender and object, a  high or low degree of similarity can unbalance the score.  Also, the false positive object information by the visual classifier influences the score negatively which lead to wrong bias estimation. "]}
{"annotations": [{"id": 86960964, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86973106, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973991, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_25", "final_parag": "We rely upon an existing range of well-known publicly available caption datasets crawled from the web and annotated by humans that assume a binary conceptualization of gender. Therefore, it is important to acknowledge that within the scope of this work, we are treating gender as strictly binary (\\ie man and woman) oversimplifies a complex and multifaceted aspect of human identity. Gender is better understood as a spectrum, with many variations beyond just two categories, and should be addressed in future work. Since all models are being trained on these datasets, we anticipate all models contain other biases (racial, cultural, \\etc). For example, an observation in Table \\ref{tb:glove}, when we remove gender bias, we notice the emergence of another bias, such as racial bias. For example, vectors representing \\textit{Black} person or women are closer together than those representing other colors, like \\textit{white}. Moreover, there is another form of bias that has received limited attention in the literature, the propagation of gender and racial bias via RL (\\eg RL-CLIPS). For instance, in Figure \\ref{fig:main_figure} the model associates gender with race, as “asian woman”.", "candidate_comments": ["We rely upon an existing range of well-known publicly available caption datasets crawled from the web and annotated by humans that assume a binary conceptualization of gender. Therefore, it is important to acknowledge that within the scope of this work, we treating gender as strictly binary (\\ie man and woman) oversimplifies a complex and multifaceted aspect of human identity. Gender is better understood as a spectrum, with many variations beyond just two categories, and should be addressed in future work. Since all models are being trained on these datasets, we anticipate all models contain other biases (racial, cultural, \\etc). For example, an observation in Table  \\ref{tb:glove}, when we remove gender bias, we notice the emergence of another bias, such as racial bias. For example, vectors representing \\textit{Black} person or women are closer together than those representing other colors, like \\textit{white}. Moreover,  there is another bias that is not well studied in the literature, the propagation of gender and racial bias via RL (\\eg, RL-CLIPS). For instance, in Figure \\ref{fig:main_figure} the model refers to the gender by race as “asian woman”.  "]}
{"annotations": [{"id": 86972379, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture", "Réécriture"]}, {"id": 86983040, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture", "Réécriture"]}, {"id": 86984399, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_26", "final_parag": "{\\textbf{If the goal is to evaluate gender bias why using pre-trained models.}} We believe this is representative of real-world scenarios where we can measure the gender bias from real data and not from a closed domain as in the human-annotated COCO Captions dataset. Moreover, in contrast to recent previous work \\cite{hirota2022quantifying}, which heavily depends on a costly human annotation dataset \\cite{zhao2021understanding}, limiting the scalability across various domains, our Gender Score is not bounded by any dataset and can be adapted for any task, including the additionally annotated existing caption dataset for the gender bias analysis.", "candidate_comments": ["We believe this is a reflection of real-world situations where we can gauge gender bias using real data rather than being confined to a closed dataset like the human-annotated COCO Captions. In contrast to recent studies (Zhao et al., 2021; Hirota et al., 2022) that heavily rely on costly human annotation datasets (Zhao et al., 2021), limiting their scalability across various tasks, our Gender Score stands out. It is not constrained by any specific dataset and can be tailored for diverse tasks, including those using additionally annotated caption datasets to address gender bias.", "Moreover, in contrast to recent studies such as those by Zhao et al. (2021) and Hirota et al. (2022), which rely heavily on expensive human-annotated datasets, making scalability across various tasks challenging, our Gender Score is not confined to a specific dataset. It can be flexibly adapted for a wide range of tasks, including those that utilize additionally annotated caption datasets for addressing gender bias \\cite{zhao2021understanding}.", "Furthermore, unlike recent studies by Zhao et al. (2021) and Hirota et al. (2022), which heavily depend on costly human annotation, making it difficult to scale for different tasks, our Gender Score is not limited by any specific dataset. It can be adapted for various tasks, including those involving additional annotated caption datasets for addressing gender bias."]}
{"annotations": [{"id": 86971577, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86973218, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974680, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_27", "final_parag": "Table \\ref{tb:benh-additional} shows a comparison result between our Gender Score and BERT$_{\\text{L}}$ /GPT-2$_{\\text{L}}$ as baselines using only the word surrounding context without any visual bias. The result shows that our gender estimation model is similar to BERT in some cases as shown in Figures \\ref{tb-fig:3}|\\ref{tb-fig:4}, which are only influenced by a strong bias. However, when the generated caption is more/less diverse, \\ie due to less/more context to predict the gender, BERT \\texttt{Mask} LM struggles with predicting the correct biased gender. In contrast to BERT, our approach relies on additional context extracted from the image and uses a language model and similarity score to measure the degree of gender-based bias prediction.", "candidate_comments": ["Table \\ref{tb:benh-additional} shows a comparison result between our score and BERT/GPT-2 as a baseline, and without any visual context. The result shows that our gender estimation model is similar to BERT in most cases, which is only influenced by a strong bias, and unlike the cosine distance score and current object-gender freq count approaches."]}
{"annotations": [{"id": 86952704, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86959186, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86965129, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_28", "final_parag": "{\\bf{Transformer \\cite{Ashish:17}.}} We use a transformer model that is trained on bottom-up and top-down features \\cite{anderson2018bottom} with cross-entropy loss with learning rate $3e-04$.", "candidate_comments": ["{\\bf{Transformer \\cite{Ashish:17}.}} We train the transformer from scratch with the bottom-up and top-down features  \\cite{anderson2018bottom} with cross-entropy loss  with different runs (the best reported  result a learning rate $3e-04$)."]}
{"annotations": [{"id": 86972277, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86980173, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86980644, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_29", "final_parag": "{\\bf{AoANet \\cite{huang2019attention}}.} We utilize the model as out-of-the-box with the best-reported result by the authors.", "candidate_comments": ["{\\bf{Transformer \\cite{Ashish:17}.}} We train the transformer from scratch with the bottom-up and top-down features  \\cite{anderson2018bottom} with cross-entropy loss  with different runs (the best reported  result a learning rate $3e-04$)."]}
{"annotations": [{"id": 86983245, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86979543, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86981363, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19130-03cf4a9a918931f8.jsonl", "id_parag": "2310.19130-03cf4a9a918931f8_30", "final_parag": "{\\bf{OSCAR$_{\\text{Large}}$ \\cite{li2020oscar}.}} To make a fair comparison with other pre-trained vision and language models, we fine-tuned the model on the COCO-Caption dataset with cross-entropy evaluation score without RL with a learning rate $5e-06$ (15 epochs), following the same paper setup.", "candidate_comments": ["{\\bf{Transformer \\cite{Ashish:17}.}} We train the transformer from scratch with the bottom-up and top-down features  \\cite{anderson2018bottom} with cross-entropy loss  with different runs (the best reported  result a learning rate $3e-04$)."]}
{"annotations": [{"id": 86725361, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}, {"id": 86733040, "completed_by": {"id": 140156}, "result": ["Différent", "Différent"]}, {"id": 86766110, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19138-25992e60aceb7ac4.jsonl", "id_parag": "2310.19138-25992e60aceb7ac4_1", "final_parag": "\\section{List of Symbols} \\begin{itemize} \\setlength\\itemsep{2pt} \\item IC: Independent-Cascade \\item DSL: Diffusion Source Localization. \n\\item IM: Influence Maximization. \n\\item RC: Rumor-Centrality. \n\\item MLE: Maximum Likelihood Estimator. \n\\item SFT: Short-Fat-Tree algorithm. \n\\item WBND: Weight Bounded Node Degree. \n\\item BFS: Breadth First Search. \n\\item DMP: Dynamic-Message-Passing algorithm. \n\\item BP: Belief Propagation algorithm. \n\\item WLOG: Without loss of generality. \n\\item DL: Distributive-law. \n\\item LTP: Law of Total Probability. \n\\item $G$: A static directed graph with $V$ and $\\vv{E}$ as the sets of nodes and directed edges respectively. WLOG, it is assumed that whenever $(i,j)$ is in $\\vv{E}$, $(j, i)$ is also in $\\vv{E}$. $E$ denotes the set of undirected edges generated by $\\vv{E}$. Each connected-component of $G$ is assumed to contain at least two nodes. \n\\item $\\partial{i}$: Set of neighbors of node $i$ in graph $G$, i.e., $\\{ k \\in V: \\{k, i\\} \\in E \\} $. \n\\item $\\mcl{C}$: Set of connected-components of $G$. Each $C \\in \\mcl{C}$ is assumed to have at least two nodes. \n\\item $I$ and $J$: Typically used to denote diffusion processes $A$ and $B$. \n\\item $\\bsl{\\lambda}_{i\\ra j}$: $ \\l\\{ \\lambda_{i\\ra j}^{A \\ra \\emptyset}, \\lambda_{i\\ra j}^{A \\ra \\{B\\}}, \\lambda_{i\\ra j}^{B \\ra \\emptyset}, \\lambda_{i\\ra j}^{B \\ra \\{A\\}} \\r\\}$. \n\\item $\\bsl{\\lambda}$: $\\l( \\bsl{\\lambda}_{i\\ra j}: (i,j)\\in \\vv{E} \\r)$. \n\\item $X_i^{(t)}$: State of node $i$ at time $t$. Also, see \\eqref{eq:xit}. \n\\item $\\mcl{X} $: $ \\{ \\emptyset, \\{A\\}, \\{B\\}, \\{A, B\\} \\}$. \n\\item $ \\mbf{X}^{(t)}$: State of network at time $t$, i.e., $\\l( X_i^{(t)} : t \\ge 0 \\r) $. \n\\item $D_{i\\ra j}^{I\\ra \\mcl{J}}$: Activation-time variable - time taken by node $i$ to infect node $j$ with process $I$ when node $j$ has state $\\mcl{J} \\in \\{ \\emptyset, \\{A,B\\}\\setminus\\{I\\} \\}$ at the time of infection attempt. \n\\item $f_{D_{i\\ra j}^{I\\ra \\mcl{J}} }$: See \\eqref{eq:fdijij}. \n\\item $\\mbf{D}_{i\\ra j}$: $\\l\\{ D_{i\\ra j}^{A \\ra \\emptyset}, D_{i\\ra j}^{A \\ra \\{B\\}}, D_{i\\ra j}^{B \\ra \\emptyset}, D_{i\\ra j}^{B \\ra \\{A\\}} \\r\\}$. \n\\item $\\mbf{D}$: $\\l( \\mbf{D}_{i\\ra j} :(i,j)\\in \\vv{E} \\r)$. \n\\item $W$: Observation-time. \n\\item $\\mcl{O}_n$: Observation-event $\\{ \\wt{\\mbf{X}}^{(W)} = \\wt{\\mbf{x}} \\} $. The subscript $n$ indicates that the observation is noisy. \n\\item $f_{\\wt{X}_i| X_i} $: Probability of observing node $i$ in state $\\wt{x}_i$ given its (true) state is $x_i$. \n\\item $f_{\\wt{\\mbf{X}} | \\mbf{X}} $: See \\eqref{eq:fo}. \n\\item $f_W$: See \\eqref{eq:fw} \\item $\\mcl{W}$: $\\{ w_{min}, w_{min}+1, \\dots, w_{max} \\} $. \n\\item $f_{\\mbf{X}^{(0)} } $: Distribution of network's initial-state. See \\eqref{eq:fx0}. \n\\item $f_{\\mbf{D}_{i\\ra j}} \\l( \\mbf{d}_{i\\ra j} \\r)$: See \\eqref{eq:fdij}. \n\\item $T_i^I$: Time at which node $i$ gets infected with process $I$. See \\eqref{eq:tii} and \\eqref{eq:tii2}. \n\\item $\\mcl{T}$: $\\{0,1,\\dots, |V|-1, \\infty \\} $. \n\\item $\\mbf{T}_i$: $ \\l\\{T_i^A, T_i^B\\r\\}$. \n\\item $\\mbf{T}$: $\\l(\\mbf{T}_i:i\\in V\\r)$. \n\\item $\\gamma_i^p $: See \\eqref{eq:gammaip}. \\item $T_{k\\ra i}^I$: Propagation-time of process $I$ from $k$ to (its neighbor) $i$. See \\eqref{eq:tkii}. \n\\item $f_{\\mbf{T}_i} $: See \\eqref{eq:fti} and \\eqref{eq:fti2}. \n\\item $\\gamma_i$: See \\eqref{eq:po}. \n\\item $\\sigma(b_i; a)$: See \\eqref{eq:sigmadef}. \n\\item $\\zeta_i$: See \\eqref{eq:zetai}. \n\\item $\\psi_{\\ra i} $: See \\eqref{eq:psii}. \n\\item $\\psi_{\\ra i}^I $: See \\eqref{eq:psiii1} and \\eqref{eq:psiii2}. \n\\item $S_{i\\ra j}^I$: See \\eqref{eq:siji}. \n\\item $\\mcl{S}$: $ \\{ 0, 1\\}$. \n\\item $\\mbf{S}_{i\\ra j}$: $\\{S_{i\\ra j}^A, S_{i\\ra j}^B\\}$. \n\\item $\\mbf{S}$: $\\l(\\mbf{S}_{i\\ra j} : (i,j)\\in \\vv{E} \\r)$. \n\\item $\\psi_{i\\xlrarrow{} j} $: See \\eqref{eq:psiijji}. \n\\item $\\psi_{i\\ra j} $: See \\eqref{eq:psiij}. \n\\item $\\psi_{i\\ra j}^I$: See \\eqref{eq:psiiji}. \n\\item $\\udl{f}_W$: See \\eqref{eq:fw1}. \n\\item $\\udl{\\gamma}_i $: See \\eqref{eq:gammai1}. \n\\item $\\Gamma$: See \\eqref{eq:biggamma}. \n\\item $f_i, f_{ij} $: See \\eqref{eq:fiandfij}. \n\\item $G_{\\Gamma}$: Inflated factor-graph representation of $\\Gamma$ given by \\eqref{eq:gbiggamma}. \n\\item $m_{f\\ra v}^{(k+1)} \\l(\\mbf{z}_{r(v)}\\r)$: See \\eqref{eq:f2v_1} and \\eqref{eq:f2v_2}. \n\\item $n_{v \\ra f}^{(k)}\\l(\\mbf{z}_{r(v)}\\r)$: See \\eqref{eq:v2f}. \n\\item $N(\\cdot)$: $N(\\cdot)$ outputs the neighborhood set of its input. \n\\item $\\mbf{m}_{f\\ra v}^{(k)}$: $\\l\\{ m_{f\\ra v}^{(k)} \\l(\\mbf{z}_{r(v)}\\r) \\r\\}_{\\mbf{z}_{r(v)}}$. \n\\item $\\mbf{n}_{v\\ra f}^{(k)}$: $\\l\\{ n_{v\\ra f}^{(k)} \\l(\\mbf{z}_{r(v)}\\r)\\r\\}_{\\mbf{z}_{r(v)}}$. \n\\item $q_{r(v)}^{(k)}\\l( \\mbf{z}_{r(v)} \\r)$: See \\eqref{eq:qv}. \n\\item $ T_{max}$: Maximum finite time for either process to reach a given node. Introduced in Assumption \\ref{assump:tmax}. \n\\item $\\mcl{T}' $: $ \\{0,1,\\dots, T_{max}, \\infty \\}$. See Assumption \\ref{assump:tmax}. \n\\item $\\pr'$: Probability measure for the (modified) diffusion-model of Assumption \\ref{assump:tmax}. \n\\item $f'_W$: See \\eqref{eq:fw2}. \n\\item $\\gamma'_i $: See \\eqref{eq:gammai2}. \n\\item $ \\Gamma'$: See \\eqref{eq:biggamma_2}. \n\\item $f'_i, f'_{ij} $: See \\eqref{eq:fiandfij_2}. \n\\item $G_{\\Gamma'}$: Inflated factor-graph representation of $\\Gamma'$ given by \\eqref{eq:gbiggamma_2}.", "candidate_comments": ["diffusion trace estimation \\item $ Y^I(t)$: Set of nodes that get infected with process $I$ at time $t$, i.e., $\\{i \\in V: T_i^I = t \\}$ \\item $\\mbf{Y}$: $\\l\\{ Y^I(t) : t \\ge 0 \\r\\}$. \n\\item $\\partial\\l( \\udl{V}\\r)$: See \\eqref{eq:partialv}. \n\\item $Z^I(t)$: See \\eqref{eq:zit}. \n\\item $\\phi_{j\\ra i}^I(\\udl{V})$. See \\eqref{eq:phijii}.", "\\item $q_{\\mbf{x}^{(0)}}\\l(\\mbf{x}^{(0)}\\r), q_{{x}_i^{(0)}}\\l({x}_i^{(0)}\\r), q_w(w), q_{{t}_i^I}(t_i^I)$: See \\eqref{eq:posteriorbeliefs}."]}
{"annotations": [{"id": 86736600, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86900896, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86900964, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19138-25992e60aceb7ac4.jsonl", "id_parag": "2310.19138-25992e60aceb7ac4_2", "final_parag": "\\item $ \\mcl{I}_1$: See \\eqref{eq:i1}. \n\\item $ \\mcl{I}_2$: See \\eqref{eq:i2}. \n\\item $ \\mcl{I}_3$: See \\eqref{eq:i3}. \n\\item $\\udl{\\mcl{T}}_i^I$: Some subset of $\\mcl{T}$. \n\\item $\\udl{\\mcl{S}}_{i\\ra j} $: Some subset of $\\mcl{S} $. \n\\item $\\udl{\\mcl{W}} $: Some subset of $\\mcl{W}$. \n\\item $V_p$: Set of all protector nodes. See Section \\ref{sec:disc:anticontagion}. \n\\item $T_{delay}^B$: Random Delay in the initiation of process $B$. \n\\item $f_{T_{delay}^B}$: See \\eqref{eq:fh}. \n\\item $\\mcl{T}_{delay}^B$: $\\l\\{h_{min}, h_{min}+1, \\dots, h_{max}, \\infty \\r\\}$.", "candidate_comments": ["diffusion trace estimation \\item $ Y^I(t)$: Set of nodes that get infected with process $I$ at time $t$, i.e., $\\{i \\in V: T_i^I = t \\}$ \\item $\\mbf{Y}$: $\\l\\{ Y^I(t) : t \\ge 0 \\r\\}$. \n\\item $\\partial\\l( \\udl{V}\\r)$: See \\eqref{eq:partialv}. \n\\item $Z^I(t)$: See \\eqref{eq:zit}. \n\\item $\\phi_{j\\ra i}^I(\\udl{V})$. See \\eqref{eq:phijii}."]}
{"annotations": [{"id": 86982141, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86982278, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86982692, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19138-aa33d080dfa3495a.jsonl", "id_parag": "2310.19138-aa33d080dfa3495a_1", "final_parag": "\\newtheorem{thm}{Theorem} \\newtheorem{prop}{Proposition} \\newtheorem{dfn}{Definition} \\newtheorem{rem}{Remark} \\newtheorem{lem}[thm]{Lemma} \\newtheorem{cor}{Corollary}[thm] \\newtheorem{fact}{Fact}", "candidate_comments": ["\\theoremstyle{remark} \\newtheorem{remark}{Remark} \\newtheorem{rems}{Remarks} \\newtheorem{para}{} \\newtheorem{summary}{Summary} -------------------------------------------------------------------"]}
{"annotations": [{"id": 86953081, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86959365, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86960212, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-088bc6cd02ec13ff.jsonl", "id_parag": "2310.19139-088bc6cd02ec13ff_1", "final_parag": "Various benchmarks, like the popular Defects4J, assist in evaluating fault localization techniques. It contains real software bugs, relevant fault-triggering, and bug fixes. However, the fault-triggering tests in Defects4J might be extracted from a version of the system where the bug is already fixed. Thus, these tests may offer clues to fault localization techniques regarding the bug's location. \nWe present a case study that examines developer knowledge in fault-triggering tests. Our results show that 77\\% of the fault-triggering tests in Defects4J contain developer knowledge. Our manual study highlights that developers often add new tests/assertions to replicate the bug or modify existing tests for regression testing, leading to the existence of such knowledge. \nWe also conduct an experiment on the impact of tests with developer knowledge on the performance of state-of-the-art spectrum-based fault localization techniques. Our experimental results show that when evaluating tests without developer knowledge, the performance of fault localization techniques significantly degrades. This is because the fault-triggering tests, which no longer contain developer knowledge, do not fail and do not contribute to locating the fault. \nFinally, we provide some potential research directions that may be done using our dataset. \nIn our study, we highlight important findings on the developer knowledge that may exist in fault-triggering tests, overestimating the performance of fault localization research in a realistic setting. We present our dataset publicly available that may be used for new or extension to existing benchmarks.", "candidate_comments": ["In particular, various benchmarks have been proposed to assist developers in evaluating new fault localization techniques. The Defects4J benchmark is a popular benchmark for fault localization."]}
{"annotations": [{"id": 86961963, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972514, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973831, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-11a3fee56c18a8d3.jsonl", "id_parag": "2310.19139-11a3fee56c18a8d3_1", "final_parag": "Developers may modify tests to change the logic of the tests when trying to fix a bug. Typically, these modifications happen when developers are trying to replicate the bug in a test. Examples of modifications include simplifying complex logic and reorganizing the code structure, which can alter the execution of the test. In Figure~\\ref{lst:testlogic}, we show an example of a test modification in this category based on the {\\sf JsonAdapterNullSafeTest} from issue \\#800 in GSON. During the bug-fixing process, developers discussed providing a ``simpler'' test to reproduce the bug. They incorporated new changes that simplified the initialization of the {\\sf Device} class from the test and removed logic that was irrelevant in triggering the fault. This contributed to improving the quality and maintenance of the test.", "candidate_comments": ["Developers often refactor tests to make them easier to understand and maintain. By improving the readability of tests, the codebase can be better understood, resulting in better performance in future releases. A readability-focused change may include simplifying complex logic, renaming variables more descriptively, or reorganizing the code structure. A good example of this category is the refactoring of the {\\sf testNullSafeBugDeserialize} method in the {\\sf JsonAdapterNullSafeTest} from a project GSON bug report. Among the changes made to the test was the use of readable JSON input, simplifying the Device class, and streamlining the implementation of {\\sf JsonAdapterFactory}. Consequently, the test becomes easier to understand and maintain, which contributes to the overall quality and stability of the codebase."]}
{"annotations": [{"id": 87251784, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87259799, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87267787, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19139-173e42c75322f2a4.jsonl", "id_parag": "2310.19139-173e42c75322f2a4_1", "final_parag": "A major advantage of Defects4J is its controlled environment that allows researchers to reproduce the bugs easily. However, as we found, 55\\% of the tests provided by Defects4J were not available when the bugs were first reported. \nHence, it is yet unknown how fault localization techniques generalize to less-controlled settings. Practically, developers may use fault localization tools to gain insights into the debugging locations before they start to fix the bugs.", "candidate_comments": ["Therefore, in this paper, we study the fault-triggering tests and its impact on fault localization. We characterize the modifications and creations of fault-triggering tests with regards to the bug report. Our analysis shows that over 77\\% of the tests were either modified or created after the bug was reported, introducing developer knowledge of the bugs."]}
{"annotations": [{"id": 86899123, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86902059, "completed_by": {"id": 62306}, "result": ["Réécriture", "Différent"]}, {"id": 86906720, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}], "agreement": 55.55555555555555, "source_file": "2310.19139-173e42c75322f2a4.jsonl", "id_parag": "2310.19139-173e42c75322f2a4_2", "final_parag": "The main contributions of this paper are: \\begin{itemize} \\item We present a case study examining the prevalence of developer knowledge in Defects4J's fault-triggering tests. Our findings show that the majority of fault-triggering tests (77\\%) contain developer knowledge of the bug.", "candidate_comments": ["Therefore, in this paper, we study the fault-triggering tests and its impact on fault localization. We characterize the modifications and creations of fault-triggering tests with regards to the bug report. Our analysis shows that over 77\\% of the tests were either modified or created after the bug was reported, introducing developer knowledge of the bugs.", "Our experimental results show that the state-of-the-art spectrum-based fault localization techniques perform significantly worse in the absence of developer knowledge in the tests. In a more realistic scenario, fault localization is carried out on the version of the system where the bug has recently been reported. In this case, the fault-triggering tests, before they are modified by developers, can not assist in capturing the fault (i.e., do not fail)."]}
{"annotations": [{"id": 87251835, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87259011, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87267917, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19139-173e42c75322f2a4.jsonl", "id_parag": "2310.19139-173e42c75322f2a4_3", "final_parag": "\\item We provide a discussion on the implications of our findings and highlight future research direction. We also made our dataset publicly available online~\\cite{our_repo}, which includes a comprehensive list of the bugs that contain developer knowledge, the details of our manual study results, and the test result and coverage that we collected for the buggy commit (before the bug was reported).", "candidate_comments": ["\\item We have made our dataset publicly available online~\\cite{our_repo}, which includes a comprehensive list of the bugs impacted by developer knowledge from Defects4J, along with the details of our manual studies."]}
{"annotations": [{"id": 86736452, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86900110, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86900580, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-46c95709eb8e56d6.jsonl", "id_parag": "2310.19139-46c95709eb8e56d6_1", "final_parag": "However, we adopted the manual analysis approaches used in prior studies~\\cite{li2019dlfinder, lamothe2020apis, ding2020towards, li2020shall} to mitigate subjectivity. Three phrases were used and two authors independently were involved in the analysis. The analysis results achieve a Cohen’s Kappa of 0.86, which indicates a substantial level of agreement~\\cite{cohen1960coefficient}. In RQ1, we also manually reviewed the relevance of the test modification to a bug. However, the main work in extracting events in relation to bug resolution was automatic and we only manually check whether any errors in the generated results.", "candidate_comments": ["However, we adopted the manual analysis approaches used in prior studies~\\cite{li2019dlfinder, lamothe2020apis, ding2020towards, li2020shall} to mitigate subjectivity. Three phrases were used and two authors independently were involved in the analysis. The analysis results achieve a Cohen’s Kappa of 0.86, which indicates a substantial level of agreement~\\cite{cohen1960coefficient}. "]}
{"annotations": [{"id": 86862490, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86901970, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86903504, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-46c95709eb8e56d6.jsonl", "id_parag": "2310.19139-46c95709eb8e56d6_2", "final_parag": "{\\bf External Validity.} Threats to external validity concern the degree to which our findings can be generalized. While our study primarily centers on the Defects4J dataset, it is crucial to recognize that other datasets with test cases and bug information might yield different outcomes. The Defects4J dataset, being a predominant benchmark for significant software engineering tasks like fault localization and automated program repair, has deeply ingrained its relevance in the research community. Numerous studies, particularly in the APR domain, lean heavily on the Defects4J dataset for evaluating Java programs. Given its widespread use, it is conceivable that some characteristics of Defects4J might have influenced research results or methodologies. However, certain key insights from our analysis, such as the impact of developer knowledge in fault-triggering tests and the observed modification categories, are not inherently tied to Defects4J. Thus, they could potentially manifest in other datasets as well. For example, datasets derived from projects with similar development practices or that undergo similar processes of curation could exhibit comparable biases or characteristics. Furthermore, it's worth mentioning that while the core of our findings is rooted in Defects4J, the underlying themes—like the influence of developer knowledge—could be pervasive in other datasets. As such, researchers exploring datasets akin to Defects4J should consider the potential implications and biases our study unveils.", "candidate_comments": ["{\\bf External Validity.} Threats to external validity are related to the generalization of our results. In this study, we mainly focus on the Defects4J dataset. Conducting the study on different datasets with test cases and bug information may have different results. \\nakhla{updated}Nevertheless, given the general nature of some of our findings, it's plausible that they could be applicable to other datasets as well. \n%For example, the analysis of new datasets can bring in new categories of modifications The Defects4J dataset is the most used benchmark in important software engineering tasks, such as fault localization and automated program repair. Quite a large body of literature, e.g., in APR, heavily relies on the Defects4J dataset for the evaluation of Java programs. The discussions, such as developer knowledge in fault-triggering tests and the identified categories of modifications on fault-triggering, derived from the analysis of Defects4J in this paper could be beneficial to future research, such as evaluating the research in the settings of no developer knowledge that is known in the tests."]}
{"annotations": [{"id": 86725934, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86899684, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86900324, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19139-66c30bd111f2a070.jsonl", "id_parag": "2310.19139-66c30bd111f2a070_1", "final_parag": "\\begin{abstract} For software testing research, Defects4J stands out as the primary benchmark dataset, offering a controlled environment to study real bugs from prominent open-source systems. However, prior research indicates that Defects4J might include tests added post-bug report, embedding developer knowledge and affecting fault localization efficacy. In this paper, we examine Defects4J's fault-triggering tests, emphasizing the implications of developer knowledge of SBFL techniques. We study the timelines of changes made to these tests concerning bug report creation. Then, we study the effectiveness of SBFL techniques without developer knowledge in the tests. We found that 1) 55\\% of the fault-triggering tests were newly added to replicate the bug or to test for regression; 2) 22\\% of the fault-triggering tests were modified after the bug reports were created, containing developer knowledge of the bug; 3) developers often modify the tests to include new assertions or change the test code to reflect the changes in the source code; and 4) the performance of SBFL techniques degrades significantly (up to --415\\% for Mean First Rank) when evaluated on the bugs without developer knowledge. We provide a dataset of bugs without developer insights, aiding future SBFL evaluations in Defects4J and informing considerations for future bug benchmarks. \n\\end{abstract}", "candidate_comments": ["The abstract is a short summary of the work to be presented in the article. \n\\begin{abstract}   To facilitate research in software testing, researchers have proposed several benchmark datasets for the community to evaluate fault localization and automated program repair techniques. Among them, Defects4J is the most widely used benchmark due to its clean and controlled environment for studying real bugs collected from popular open-source systems. When localizing faults, spectrum-based fault localization (SBFL) techniques rely mainly on analyzing the coverage of the fault-triggering tests. However, prior research suggests that Defects4J may contain some tests that were added by developers after the bug was reported or fixed. Such tests may contain developer knowledge of the bug, which compromises the effectiveness of fault localization techniques. In this paper, we conduct a comprehensive study on the fault-triggering tests in Defects4J. Our principal contribution lies in revealing the extent of developer knowledge in fault-triggering tests and their implications on SBFL techniques. We study the timelines of the changes that developers made to the fault-triggering tests with respect to bug report creation time. Then, we study the effectiveness of SBFL techniques without developer knowledge in the tests. We found that 1) 55\\% of the fault-triggering tests were newly added to replicate the bug or to test for regression; 2) 22\\% of the fault-triggering tests were modified after the bug reports were created, containing developer knowledge of the bug; 3) developers often modify the tests to include new assertions or change the test code to reflect the changes in the source code; and 4) the performance of SBFL techniques degrades significantly (up to --415\\% for Mean First Rank) when evaluated on the bugs without developer knowledge. We provide a dataset of bugs without developers' knowledge in tests, which can help future studies conduct in-depth evaluations of their SBFL techniques in different settings when using Defects4J. Our findings also provide insight into the development activities that need to be considered when creating future bug benchmarks. \n\\end{abstract}"]}
{"annotations": [{"id": 86985889, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86985824, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86985341, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19139-66c30bd111f2a070.jsonl", "id_parag": "2310.19139-66c30bd111f2a070_2", "final_parag": "\\begin{CCSXML} <ccs2012> <concept> <concept_id>10011007.10011074</concept_id> <concept_desc>Software and its engineering~Software creation and management</concept_desc> <concept_significance>500</concept_significance> </concept> <concept> <concept_id>10011007.10011074.10011099.10011102.10011103</concept_id> <concept_desc>Software and its engineering~Software testing and debugging</concept_desc> <concept_significance>500</concept_significance> </concept> </ccs2012> \\end{CCSXML}", "candidate_comments": ["\\begin{CCSXML} <ccs2012>    <concept>        <concept_id>10011007.10011074.10011099.10011102.10011103</concept_id>        <concept_desc>Software and its engineering~Software testing and debugging</concept_desc>        <concept_significance>500</concept_significance>        </concept>  </ccs2012> \\end{CCSXML}"]}
{"annotations": [{"id": 86864510, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86902152, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86903550, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-87860d74e2f93fd5.jsonl", "id_parag": "2310.19139-87860d74e2f93fd5_1", "final_parag": "\\phead{Discussion.} Since the effectiveness of fault localization techniques degrades significantly, we further investigate the possible causes, other than that the tests contain developer knowledge of the bug. We examine how many bugs whose fault-triggering tests (from Pattern 3 and 4) actually failed in {\\em vBuggy} (i.e., the version when the bug was first reported). Note that our study focuses on 263 out of 305 bugs, as some earlier versions of the system could not be retrieved. Table~\\ref{tab:nonfailing} shows the number of bugs with failed fault-triggering tests in {\\em vBuggy}. We noticed that in both Pattern 3 and Pattern 4, a small percentage (15\\%) of fault-triggering tests did not cause any failure in {\\em vBuggy}, indicating that they were unable to detect the fault. \nOn the other hand, the remaining fault-triggering tests (85\\%) failed and assisted in fault localization. \nThis result suggests that the majority of these tests did fail in {\\em vBuggy}, but they were not effective in detecting the bugs.", "candidate_comments": ["On the other hand, executing fault-triggering tests in Pattern 3 without developer knowledge results in the tests failing to trigger the bugs or producing any failures, despite the bugs still being present in the system. "]}
{"annotations": [{"id": 86721315, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}, {"id": 86901549, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}, {"id": 86901793, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19139-87860d74e2f93fd5.jsonl", "id_parag": "2310.19139-87860d74e2f93fd5_2", "final_parag": "Our finding shows that developers’ knowledge may not always be directly incorporated into the fault-triggering tests themselves. Future research may explore bugs from Pattern 4 to investigate and characterize the instances where developers’ knowledge is introduced outside of the fault-triggering tests. \nNevertheless, our findings reveal that fault-triggering tests in Pattern 4, which were not modified by developers during the bug fixing process, resulted in better performance for fault localization, as indicated by both MFR and MAR metrics. \nThese bugs are better suited for evaluating fault localization techniques in practice.", "candidate_comments": ["On the other hand, executing fault-triggering tests in Pattern 3 without developer knowledge results in the tests failing to trigger the bugs or producing any failures, despite the bugs still being present in the system. ", "\\peter{@Nahkla, please try to write the approach here} \\anran{we have already collected the test results} Table~\\ref{tab:nonfailing} show the number of bugs with and without failed fault-triggering tests in {\\em vBuggy}. A majority of the bugs had fault-triggering tests (i.e., 88\\%) failing in {\\em vBuggy}, while the remaining 12\\% with no failed fault-triggering tests. Therefore, for the "]}
{"annotations": [{"id": 86962200, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973306, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973724, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-87860d74e2f93fd5.jsonl", "id_parag": "2310.19139-87860d74e2f93fd5_3", "final_parag": "Future studies may benefit from using bugs/tests from Pattern 4 to assess fault localization techniques in a more realistic setting.", "candidate_comments": ["\\peter{@Anran, please complete the discussion part. Basically, we want to say pattern 3 data is not good. Pattern 4 is fine. Future studies might want to use Pattern 4 for evaluation in a more real setting}"]}
{"annotations": [{"id": 86982140, "completed_by": {"id": 140156}, "result": ["Réécriture", "Différent"]}, {"id": 86985280, "completed_by": {"id": 62306}, "result": ["Différent", "Réécriture"]}, {"id": 86985398, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent"]}], "agreement": 100.0, "source_file": "2310.19139-bce6d6dbedd2df84.jsonl", "id_parag": "2310.19139-bce6d6dbedd2df84_1", "final_parag": "Hence, our findings call for more empirical studies to further understand developers' bug-fixing activities, and how to consider such activities when creating benchmark datasets.", "candidate_comments": ["we provide a clean version of the data  \\textbf{{\\em Future studies may need to consider developers' bug-fixing process when creating benchmarks.}} ", "\\peter{check}We also found cases where developers fixed the bug first, and then create a bug report for documentation purposes. "]}
{"annotations": [{"id": 86869403, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86900382, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86900557, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-bce6d6dbedd2df84.jsonl", "id_parag": "2310.19139-bce6d6dbedd2df84_2", "final_parag": "\\phead{Future studies may need to consider developers' bug-fixing process when creating benchmarks.} In RQ2, we found that fault-triggering tests may not be available when developers receive a bug report. Developers may then develop fault-triggering tests while working on a bug fix to replicate the bug.", "candidate_comments": ["Future studies may focus on developing new techniques to remove developer knowledge from affected fault-triggering tests. "]}
{"annotations": [{"id": 86972636, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86981322, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86981269, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19139-c4572e226cd4c519.jsonl", "id_parag": "2310.19139-c4572e226cd4c519_1", "final_parag": "However, only a few studies~\\cite{koyuncu2019ifixr, liu2021critical} have pointed out the presence of future tests. \n~\\citet{koyuncu2019ifixr} proposed a bug report-driven program repair technique in their study and noted that tests that were modified during bug fixes may impact the evaluation of their technique. \nSome of the authors of this work made the same observation in their following work~\\cite{liu2021critical} when evaluating the biases in automated program repair benchmarks. However, there has been no systematic study on the prevalence of future tests, and prior studies did not evaluate the impacts of these tests on fault localization techniques.", "candidate_comments": ["However, some studies~\\cite{koyuncu2019ifixr, liu2021critical} have pointed out the presence of future tests. ~\\citet{liu2021critical} observed that fault-triggering tests may not be available when bug reports were first submitted. ~\\citet{koyuncu2019ifixr} found that some tests may be modified during bug fixes. \n~\\citet{koyuncu2019ifixr} found that some changes may be applied to fault-triggering tests on the bug fix commits of Defects4J bugs. \n\\peter{However, prior studies only ... and did not ... }"]}
{"annotations": [{"id": 86960490, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86986236, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86988223, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19139-c4572e226cd4c519.jsonl", "id_parag": "2310.19139-c4572e226cd4c519_2", "final_parag": "\\phead{Fault Localization and Program Repair Techniques.} Due to the importance and high cost of bug-fixing activities, a large number of research studies focus on fault localization (FL)~\\cite{b2016learning, TraPT, FLUCCS, MULTRIC, zheng2016fault, DeepFL} and automated program repair (APR)~\\cite{icse20,li2022dear,saha2017elixir}. The Defects4J dataset has been used extensively as a benchmark in the evaluation of FL and APR techniques.", "candidate_comments": ["There are some papers that say there are developer knowledge in the bugs, but they are never discussed. What are the new things that we are providing and what is missing in those papers. \nWe can briefly discuss the bug datasets and say some may have the same issues as defects4j, and discuss t-evos as an alternative. \nWe can have a part of FL and APR, but it should not be the main focus. \n\\peter{this part should be expanded, and also mention other benchmarks}\\phead{Using Defects4J for Evaluating Fault Localization and Program Repair Techniques.} Due to the importance and high cost of bug-fixing activities, a large number of research studies focus on fault localization (FL)~\\cite{b2016learning, TraPT, FLUCCS, MULTRIC, zheng2016fault, DeepFL} and automated program repair (APR)~\\cite{icse20,li2022dear,saha2017elixir}. The Defects4J dataset has been used extensively as a benchmark in the evaluation of FL and APR techniques. \n\\peter{discuss briefly what they do and how they are related to defects4j}"]}
{"annotations": [{"id": 86971447, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86973527, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974817, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19139-c4572e226cd4c519.jsonl", "id_parag": "2310.19139-c4572e226cd4c519_3", "final_parag": "Orthogonal to coverage utilization, deep learning-based approaches~\\cite{lou2021boosting, DeepFL, cnn-fl, zhang2019cnn, icse21-fl} learn embeddings of test coverage information to generate the suspicious score of code elements. For example, GRACE~\\cite{lou2021boosting} proposes a novel graph representation that combines method and test coverage information, and learns to rank the faulty methods.", "candidate_comments": ["DeepFL~\\cite{DeepFL} was shown to improve the method-level FL approach TraPT~\\cite{TraPT}. CNN-FL~\\cite{cnn-fl} takes the test coverage matrix with passing/failing information into a CNN model, then predicts the suspicious elements.  DeepRL4FL~\\cite{icse21-fl} learns and combines the embeddings from test coverage matrix, source code, and execution information to identify the suspicous code elements.  GRACE~\\cite{lou2021boosting} proposes a new graph representation for a method and test coverage information and learns to rank the faulty methods.  Fixlocator~\\cite{li2022fault} also learns to vectorize the test coverage information and failing stack traces and aims to locate multiple co-fixing statements in a fix for a fault. "]}
{"annotations": [{"id": 86904151, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86964670, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86965080, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-c4572e226cd4c519.jsonl", "id_parag": "2310.19139-c4572e226cd4c519_4", "final_parag": "All of the above mentioned techniques run the test cases and source code of Defects4J to collect the test coverage information and utilize different techniques to calculate the suspiciousness scores for the relevant code elements.", "candidate_comments": ["Defects4J dataset provides a clear mapping from source code to test cases, thus it has become the most widely used dataset in evaluating the proposed relevant techniques on Java programs. "]}
{"annotations": [{"id": 86962887, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86975120, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86976777, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-c4572e226cd4c519.jsonl", "id_parag": "2310.19139-c4572e226cd4c519_5", "final_parag": "In this work, our objective is to evaluate the influence of developers' knowledge on the performance of state-of-the-art SBFL techniques. This assessment can serve as a baseline for other coverage-based fault localization techniques.", "candidate_comments": ["In our paper, we conducted a systematic study of the change patterns of the tests and analyzed the existence of developers' knowledge. Additionally, we evaluated the impact of these patterns on the performance of state-of-the-art spectrum-based fault localization techniques. \nThus, selecting test cases and their timelines into analysis is very important and wrongly selected ones could bring a huge bias into the analysis of the performance evaluation."]}
{"annotations": [{"id": 86871479, "completed_by": {"id": 70661}, "result": ["Différent", "Réécriture", "Différent"]}, {"id": 86959030, "completed_by": {"id": 62471}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86959863, "completed_by": {"id": 126844}, "result": ["Différent", "Différent", "Différent"]}], "agreement": 83.33333333333334, "source_file": "2310.19139-c4572e226cd4c519.jsonl", "id_parag": "2310.19139-c4572e226cd4c519_6", "final_parag": "One common approach in the literature to evaluate the correctness of the patches is to verify that the candidate patch makes the original program pass all test cases. Our work investigates the prevalence of developers' knowledge in tests, which may impact prior APR research that was performed on Defects4J. \nIn addition, fault localization is a crucial step in the APR research. It helps to reduce the search space for generating patches. The test cases, particularly the fault-triggering tests that cover the buggy locations, play a crucial role in determining which part of the code should be repaired.", "candidate_comments": ["In our paper, we conducted a systematic study of the change patterns of the tests and analyzed the existence of developers' knowledge. Additionally, we evaluated the impact of these patterns on the performance of state-of-the-art spectrum-based fault localization techniques. \nThus, selecting test cases and their timelines into analysis is very important and wrongly selected ones could bring a huge bias into the analysis of the performance evaluation.", "\\anran{} how ARP uses d4j, how ARP may be affected Defects4J is also extensively and widely used in the evaluation of APR techniques. \nThe primary objective of these techniques is to generate patches that fix the bug. One common approach in the literature to evaluate the correctness of the patches is to verify that the candidate patch makes the original program pass all test cases.  For instance, ~\\citet{weimer2009automatically, le2011genprog} proposes to generate patches through genetic operations of mutation and crossover. Other studies~\\cite{liu2018mining,liu2018closer,koyuncu2020fixminer} generate repair actions from past fix patterns. Around the same time, learning-based approaches ~\\citet{} have been widely explored in the APR literature.", "One common approach in the literature to evaluate the correctness of the patches is to verify that the candidate patch makes the original program pass all test cases. "]}
{"annotations": [{"id": 86962430, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86975326, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86975447, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-c4572e226cd4c519.jsonl", "id_parag": "2310.19139-c4572e226cd4c519_7", "final_parag": "Some prior studies~\\cite{yang2017better,xiong2017precise} also use existing test cases to eliminate overfitted patches and improve the quality of generated patches, which may also be impacted by future tests.", "candidate_comments": ["have also proposed to extract change patterns from the failing tests to improve patch generations. These techniques are impacted by future tests."]}
{"annotations": [{"id": 86987142, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86972441, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86989518, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19139-da76cfcaec18606d.jsonl", "id_parag": "2310.19139-da76cfcaec18606d_1", "final_parag": "We did not consider the system Chart from Defects4J since it does not use Git as the version control system, and we rely on analyzing the development history in Git repositories to understand the changes on the fault-triggering tests for our study. \nIn total, we conducted our experiments on 809 bugs from 16 systems. Note that, one bug may have more than one fault-triggering test, so there are more fault-triggering tests than the number of bugs.", "candidate_comments": ["\\subsection{\\peter{RQ1: (we might need to find a better RQ title)What are the of change patterns in the fault triggering tests?} Studying the Characteristics of Change Patterns.}"]}
{"annotations": [{"id": 87251875, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87259522, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87267832, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19139-fdae3457e0521420.jsonl", "id_parag": "2310.19139-fdae3457e0521420_1", "final_parag": "\\phead{Motivation.} When localizing faults, SBFL techniques rely mainly on analyzing the coverage of the fault-triggering tests.", "candidate_comments": ["Prior SBFL rely on the coverage of fault-triggering tests to locate faults~\\cite{zhang2017boosting, wen2019historical, chen2022useful, xie2013metamorphic, abreu2007accuracy, abreu2006evaluation}."]}
{"annotations": [{"id": 87252011, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}, {"id": 87258711, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}, {"id": 87268285, "completed_by": {"id": 62306}, "result": ["Différent", "Réécriture"]}], "agreement": 77.77777777777777, "source_file": "2310.19139-fdae3457e0521420.jsonl", "id_parag": "2310.19139-fdae3457e0521420_2", "final_parag": "\\iuhead{Date of fault-triggering test creation and modification:} We identified all the commits that are associated with the fault-triggering tests and analyzed when the commits happened (e.g., before or after the bug was reported/fixed). We used the git command ``\\inlineCode{git log -L:[funcname]:[file]}'' to identify the list of commits that modified the fault-triggering test and the modification date.", "candidate_comments": ["We collected the date and time information provided in the output of this command to track the development activities associated with each fault-triggering test. This allowed us to better understand how the fault was identified and resolved over time with respect to the changes in fault-triggering tests. \n\\peter{maybe remove this, I don't quite get this detail}Note that if a test is inherited from a parent class, we perform our analysis directly on the parent class since any changes would only occur in that class.", "We study the commit messages and related code changes to examine if the modification adds developers' knowledge about the bug to the tests. \nWhen the timelines included test modifications during bug resolution, we ensured that they were relevant to the bugs by manually examining all the related test modifications in those timelines, referring to commit messages and related code changes to comprehend the nature of the test modifications. "]}
{"annotations": [{"id": 87252263, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87259741, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87268370, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-fdae3457e0521420.jsonl", "id_parag": "2310.19139-fdae3457e0521420_3", "final_parag": "\\iuhead{Bug fix date:} In addition to the bug report creation/resolution time, we study the time of the bug fix to understand if a test was modified before or after the fix became available. We used the ``\\inlineCode{git log [commit]}'' command to determine the exact date and time of the bug fixes. We then align the time of the bug fixes with modifications made to the fault-triggering tests.", "candidate_comments": ["We study the commit messages and related code changes to examine if the modification adds developers' knowledge about the bug to the tests. \nWhen the timelines included test modifications during bug resolution, we ensured that they were relevant to the bugs by manually examining all the related test modifications in those timelines, referring to commit messages and related code changes to comprehend the nature of the test modifications. "]}
{"annotations": [{"id": 87252560, "completed_by": {"id": 70661}, "result": ["Différent", "Différent", "Différent"]}, {"id": 87259650, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent", "Différent"]}, {"id": 87268033, "completed_by": {"id": 62306}, "result": ["Réécriture", "Différent", "Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-fdae3457e0521420.jsonl", "id_parag": "2310.19139-fdae3457e0521420_4", "final_parag": "\\iuhead{Change patterns of the tests and their relevance to the bugs:} We arranged the events -- bug report creation and resolution, creation and modification of fault-triggering tests, and bug fix time -- in chronological order to reconstruct the timeline. In particular, we focused on the creation and modifications of fault-triggering tests with respect to bug resolution. \nWe used an automated script to sequence the events into timelines. We then manually studied the commit messages and related code changes to examine if the added/modified tests include developer knowledge about the bug.", "candidate_comments": ["We study the commit messages and related code changes to examine if the modification adds developers' knowledge about the bug to the tests. \nWhen the timelines included test modifications during bug resolution, we ensured that they were relevant to the bugs by manually examining all the related test modifications in those timelines, referring to commit messages and related code changes to comprehend the nature of the test modifications. ", "We classified tests that were created after the bug report was filed as Pattern 1 and tests that were created before the bug report as Pattern 4. If the test did not exist before the bug report, we classified it as Pattern 2, and if the test existed before the bug report, we classified it as Pattern 3.", "We used an automated script to divide the tests into three groups: created before the bug report creation with no modifications during resolution, created after the bug report creation without modifications, and modified during the resolution process. The timeline for tests in the first two groups is conclusive since no additional events occurred beyond the test creation. As a result, we categorized them as Pattern 1 and Pattern 4. For the third group, we manually reviewed the modifications to confirm their relevance to the bug and accuracy in the timeline. \nWe categorized the timelines into common change patterns based on whether changes to the fault-triggering tests occurred before or during the bug report."]}
{"annotations": [{"id": 86952585, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86961455, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86960016, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19139-fdae3457e0521420.jsonl", "id_parag": "2310.19139-fdae3457e0521420_5", "final_parag": "Through manual inspection, we found that all the fault-triggering tests that belong to this pattern were added by developers as part of the bug-fixing process. Namely, nearly 50\\% of the fault-triggering tests in Defects4J were regression tests that were added deliberately to replicate/prevent the bug.", "candidate_comments": ["Through manual inspection, we identified 815 instances of this pattern, which represents nearly half of all fault-triggering tests. In all these instances, developers included new fault-triggering tests as part of the bug fix, after the creation of the bug report. \nCli 144 (Cli_14_BugCLI144Test_testFileValidator) https://issues.apache.org/jira/browse/CLI-144"]}
{"annotations": [{"id": 87252804, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87259153, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87268071, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19139-fdae3457e0521420.jsonl", "id_parag": "2310.19139-fdae3457e0521420_6", "final_parag": "\\iuhead{Pattern 3: Test modified after bug report creation (22\\%).} In practice, some bugs reported by users or other developers cannot be revealed by existing tests. Hence, when fixing these bugs, developers may modify and enhance the tests for regression testing purposes. For example, in COMPRESS-10, the test {\\sf UTF8ZipFilesTest} was enhanced as part of the bug fix. Developers modified the assertions to better capture the bug. It is possible that the test modification is not related to the bug fix. However, after our manual analysis, we found that 99\\% (357/362) of the tests contained changes that alter the test execution for replicating the bug, while the remaining 1\\% (5/362) did not introduce new knowledge to the tests (e.g., code re-styling, and enabling or disabling a failed test). In short, we find that most modifications to the fault-triggering tests were done after the bug was reported and were adding developers' knowledge of the bug in the tests.", "candidate_comments": ["However, such test modifications only happen after the bug is reported or fixed. \\peter{how many are verified that they are related to the bug?}\\anran{updated}We manually studied the tests, code changes and bug reports. We observed that 99\\% (372/377) of the tests contain changes that alter the test execution, while the remaining 1\\% (5/377) did not introduce new knowledge to the tests (e.g., code re-styling, and enabling or disabling a failed test). \n295/300"]}
{"annotations": [{"id": 87252880, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87258878, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87267713, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19155-66c30bd111f2a070.jsonl", "id_parag": "2310.19155-66c30bd111f2a070_1", "final_parag": "\\subsection{Case Study} The pilot study undertaken in this work relates to a group of residential apartments whose energy consumption is managed by a demand aggregator. Following works such as~\\cite{swiss_mpc_frequency}, we assume that the aggregator offers a power tracking service~(similar to secondary frequency control) over a short time interval~(sub-hour), when required. \nDuring normal operation (when no DR service is needed), the aggregator is responsible for maintaining the indoor room temperature in each household within the user defined comfort bounds. \nThese comfort bounds refer to a narrow band of $\\pm$ 1{\\centigrades} around a target room temperature that is set by each user. \nThis normal operation represents the business-as-usual~(BAU) scenario and the aggregator aims to maximize the comfort of the users by maintaining the room temperature as close to the target setpoint as possible.", "candidate_comments": ["This leads to an energy coordination problem where the aggregator must choose a deviation strategy that minimizes the energy deviation (and hence cost) of each household, whilst ensuring necessary user comfort. "]}
{"annotations": [{"id": 86862854, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86900862, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86900825, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19155-66c30bd111f2a070.jsonl", "id_parag": "2310.19155-66c30bd111f2a070_2", "final_parag": "However, in case of a DR event, the aggregator must switch from this BAU scenario and follow the required power signal requested by the system operator. \nTo do so, the aggregator must adjust the power consumption of different households, to ensure that the aggregate power consumption of all households together follows the requested tracking signal.", "candidate_comments": ["This leads to an energy coordination problem where the aggregator must choose a deviation strategy that minimizes the energy deviation (and hence cost) of each household, whilst ensuring necessary user comfort. "]}
{"annotations": [{"id": 86986145, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86986450, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86989822, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19155-66c30bd111f2a070.jsonl", "id_parag": "2310.19155-66c30bd111f2a070_3", "final_parag": "This section formulates the energy coordination problem from the perspective of the demand aggregator. \nFirst, we model each household as a partially observable Markov Decision Process~(MDP). MDP is a commonly used framework for modelling sequential decision-making problems and is characterized by a state space~$\\mathbf{X}$, an action space~$\\mathbf{U}$, a state transition function~$f$ and a cost function~$\\rho$~\\cite{sutton-barto}. \nIn our case, the \\emph{state} for house $h$ is defined as $\\textbf{x}^{h}_t = \\{t, T^{h}_{o,t}, T^{h}_{r,t-k}, \\ldots, T^{h}_{r,t}\\}$, \\ie the time-of-day~$t$, the past $k$ room temperature values, the current room temperature~$T^{h}_{r,t}$ and outside air temperature~$T^{h}_{o,t}$. \nAn \\emph{action}~$u^{h}_t$ is the heating power (ON or OFF), and the \\emph{cost} function is the amount of energy consumed~($G^{h}_t(\\textbf{x}^{h}_{t}, u^{h}_t)$). The transition function~($f(\\textbf{x}^{h}_{t}, u^{h}_t)$) relates to the natural evolution of room temperature over time~(duration of each time step is 15 minutes). \nNote that the actual system cannot be observed fully, with essential quantities such as building thermal mass being hidden, making it a partially observable MDP. \nTo mitigate this lack of complete information, we use a (short) history of past room temperature values over the last $k$ time steps~\\cite{sutton-barto}.", "candidate_comments": ["This leads to an energy coordination problem where the aggregator must choose a deviation strategy that minimizes the energy deviation (and hence cost) of each household, whilst ensuring necessary user comfort. "]}
{"annotations": [{"id": 86952460, "completed_by": {"id": 140156}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86958808, "completed_by": {"id": 62471}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86960711, "completed_by": {"id": 70661}, "result": ["Différent", "Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_1", "final_parag": "\\email{smr48@psu.edu} \\affiliation{ \\institution{Pennsylvania State University} \\city{State College} \\state{Pennsylvania} \\country{USA} \\postcode{16803} }", "candidate_comments": ["\\affiliation{%  \\institution{Pennsylvania State University}  \\city{State College}  \\state{Pennsylvania}  \\country{USA}}", "\\affiliation{%  \\institution{Pennsylvania State University}  \\city{State College}  \\state{Pennsylvania}  \\country{USA}}", "\\affiliation{%  \\institution{Pennsylvania State University}  \\city{State College}  \\state{Pennsylvania}  \\country{USA}}"]}
{"annotations": [{"id": 87252922, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87259964, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87267868, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_2", "final_parag": "\\begin{itemize} \\item\\textbf{RQ1}: What is the current state of reproducibility and replicability in India and to what extent do researchers in India integrate open science practices in their own work? \\item\\textbf{RQ2}: What are the experiences of researchers in India when they attempt to replicate the work of their peers? \\item\\textbf{RQ3}: What systematic incentives and changes are needed for broader integration and promotion of reproducible research practices? \\item\\textbf{RQ4}: What are the opportunities and limitations of AI-empowered tools to support reproducible and replicable research workflows, according to our participants? What features should these tools have?", "candidate_comments": ["We aim to understand the variables and factors contributing to the lack of reproducibility, identify essential features to design a reproducibility prediction tool, and best practices employed by the researchers to ensure transparency. Furthermore, we pinpoint the challenges encountered while attempting reproducible research practices, and propose potential changes to foster better research practices. Our study also addresses the strategies and design requirements for the development of Reproducibility Prediction tool."]}
{"annotations": [{"id": 87253014, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87259594, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87268334, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_3", "final_parag": "\\subsubsection{Survey recruitment} Our survey was designed to validate primary findings from our interviews across a wider community, in particular, given the multiculturalism in India and the fact that most of our interview participants were from the eastern part of the country.", "candidate_comments": ["Data for this study was collected through interviews and surveys. Data collection procedures for each are detailed as follows. % has two main sections. The first section describes the participants for the interview study and the next section describes the data collection for the survey.  "]}
{"annotations": [{"id": 86735053, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86828284, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86836704, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_4", "final_parag": "All participants were familiar with the concept of reproducibility, but only a few (P7, P8, and P12) were aware of ongoing scientific dialogue around challenges to reproducibility, i.e., the replication crisis. For example, P3 mentioned: \\quotes{I have heard about reproducibility, but not the crisis part. The majority of the researchers in my lab care about reproducibility. There are some who do not, but that percentage is rare.}{-P3}", "candidate_comments": ["Furthermore, the interviews consisted of outliers who do not regard this as a \\textit{crisis} as mentioned by  Interview participant P12, who initially lacked awareness of the \\textit{crisis}, when provided with a description, views the reproducibility issue as a problem concerning the presentation and documentation of the results rather than a \\textit{crisis}. \n\\quotes{We do not just take that as a crisis, but that is an issue regarding the presentation of results that we always admit because I belong to the measurement and instrumentation specialization in our department.}{P12}"]}
{"annotations": [{"id": 86734671, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86901094, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86901309, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_5", "final_parag": "\\quotes{The code we build must be made available to the reviewers along with very good documentation about how to run it. The datasets used for experimentation are to be made available publicly.}{-Survey Participant}", "candidate_comments": ["\\quotes{The most important thing to do, in my opinion, is to make the code public and make it well documented. And then meticulously try to address any queries that come from someone trying the code.}{-Survey Participant}"]}
{"annotations": [{"id": 86971353, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86973066, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86973566, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_6", "final_parag": "In terms of publication models, participants highlighted scarcity of venues that prioritize publishing reproducible research. P7 articulated this sentiment: \\quotes{If there was a journal of reproducibility, then people will rush to do research to publish in the journal. Right? That is one thing that is hugely necessary to increase these practices.}{-P7} Another challenge mentioned in the interviews is that open-access journals are not cost-free, which could serve as an incentive if they were not required to cover the open access fees. \n\\quotes{In some of the journals we have to pay for open access. even for some transactions. If they can make the paper free of cost to connect. This will be a great inspiration for those who are dedicated.}{-P13}", "candidate_comments": ["\\quotes{Many times the code and data can be obtained from the researchers. However, sometimes they do not respond and do not make the data/code available either. That is quite frustrating really. However, sometimes the easy way out is just to implement ourselves, compare, and then report in the paper, mentioning the code/ data was unavailable.}{-Survey Participant}"]}
{"annotations": [{"id": 87253505, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87258991, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87267825, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_7", "final_parag": "\\subsection{Perceptions of a computational tool for estimating replicability} Here we explore the perceptions of the interviewed researchers concerning a prototype AI-driven tool to estimate research replicability \\cite{rajtmajer2022synthetic}. From our interviews, three themes emerged.", "candidate_comments": ["This section describes the perception of our interview participants towards AI and the Human-AI collaboration. Also, it describes the support towards advocating research reproducibility which is highly motivating for future studies.  \\subsubsection*{Support for the tool}%Advocating Research Reproducibility} The overall response to the study of advocating reproducibility crisis was very positive and encouraging. All the professors and scholars exhibited their interest in the study. "]}
{"annotations": [{"id": 86964782, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86972470, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86974647, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_8", "final_parag": "P18 clearly mentioned the importance of data sharing. He thinks, if he doesn't share his collected data with others then it is not possible to reproduce the work of the other researchers. \n\\quotes{Say, I'm doing primary survey-based research. Okay, then, I've got the data. I have to share the data with others with my publication and reported results. Then only others can reproduce my results, of course, otherwise, without the data, how can they reproduce my results?}{-P18}", "candidate_comments": ["P6, a graduate student in machine learning mentioned the first thing that comes to his head is open code and data for evaluation. \n\\quotes{When I evaluate any paper I try to see if there is any open source code and data. So that is like the first thing that I see in a paper}{P6}"]}
{"annotations": [{"id": 87253756, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87258848, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87268320, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_9", "final_parag": "Our findings motivate the establishment of institutions tasked with promoting reproducibility, transparency, and openness within the academic ecosystem in India. Our findings also suggest that there mere existence of such institutions, without practical implementation at the grassroots level, will not effectively address the problem. They must be deeply and meaningfully embedded. Ideally, these institutions should also provide tangible assistance to researchers navigating best practices, e.g., helping them share research artifacts with proper documentation. They might also develop metrics to help quantify researchers' engagement with open science practices. These metrics might offer a means by which to shift incentives and prioritize quality over quantity of publication outputs.", "candidate_comments": ["The first thing that we are proposing is that there should be a centralized service that will promote the importance of reproducibility, transparency, and openness in every Indian university through some workshops. Just having an open science community but not implementing it at the ground level will not solve the problem. Also, they should highlight the importance of artifact sharing with proper documentation. These are the very basic qualities that researchers should have from the very beginning of their careers. The focus of these centralized services will be spreading awareness about the reproducibility crisis, open science, openness, and good research practices throughout all the universities in India. We recommend the service team should discuss with the researchers the importance of reproducibility crisis given by the peer review process and the conference protocols. They should discuss the current changes and the protocols that all the researchers need to follow for transparency and openness. The Universities should focus on the quality of the papers, not the quantity. "]}
{"annotations": [{"id": 87253833, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87259933, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87268349, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_10", "final_parag": "We suggest that publications and funding institutions consider clear guidelines for transparency, reproducibility, and openness. These criteria should be rigorously assessed by reviewers as part of the peer review process. It is important to acknowledge, however, that this may increase the workload of reviewers who currently lack sufficient incentives for such tasks. In fact, challenges to research reproducibility and challenges to peer review are deeply interrelated, and solutions will necessarily be intertwines. One might imagine a dedicated group responsible for verifying artifacts provided by authors, but again, this raises systemic and organizational questions beyond the scope of this work.", "candidate_comments": ["The second thing we recommend to all the conference chairs and the journal editors those who don't have any guidelines for transparency, reproducibility, and openness to add a clear guideline for paper submission. This will help to decrease the reproducibility crisis and will increase research ethics. Also, these submission criteria should be verified by the reviewers during the peer review process. One of the drawbacks is that this can increase their workload and they don't have any incentive for this work. There an AI-empowered tool for reproducibility prediction may be helpful for one-step verification. Also, we are suggesting that there should be one group that will verify the artifact provided by the authors."]}
{"annotations": [{"id": 87254098, "completed_by": {"id": 70661}, "result": ["Différent", "Réécriture"]}, {"id": 87259913, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent"]}, {"id": 87267993, "completed_by": {"id": 62306}, "result": ["Réécriture", "Différent"]}], "agreement": 100.0, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_11", "final_parag": "As evident from our findings, that there is notable interest among Indian researchers, in improving reproducibility and prioritizing transparency and openness. However, the lack of incentives discourages researchers from investing significant effort in integrating open science practices and strictly auditing the replicability of their work.", "candidate_comments": ["The second thing we recommend to all the conference chairs and the journal editors those who don't have any guidelines for transparency, reproducibility, and openness to add a clear guideline for paper submission. This will help to decrease the reproducibility crisis and will increase research ethics. Also, these submission criteria should be verified by the reviewers during the peer review process. One of the drawbacks is that this can increase their workload and they don't have any incentive for this work. There an AI-empowered tool for reproducibility prediction may be helpful for one-step verification. Also, we are suggesting that there should be one group that will verify the artifact provided by the authors.", "As observed from the findings there is a lack of interest in Indian researchers for reproducibility checks, transparency, and openness because no potential incentive is provided for this. Citation benefits seem to be very little for the researchers to put all these efforts into preserving quality\\cite{piwowar2013data}. Therefore we recommend different types of incentives that can reward them enough for their effective contribution. "]}
{"annotations": [{"id": 87254116, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87259212, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87268257, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_12", "final_parag": "\\subsubsection{Funding and publication} The primary incentives in academic research are publications and funding. Replication studies encounter challenges in finding suitable publication venues, as the majority of journals and conferences prioritize novelty.", "candidate_comments": ["The primary incentives in an academic research career are centered around publication and funding. Replication efforts or articles on reproducibility often face challenges in finding suitable venues, as majority of journals and conferences prioritize research novelty. This can result in lack of motivation and recognition for replication efforts. "]}
{"annotations": [{"id": 86904503, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86972746, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973879, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_13", "final_parag": "To address this issue, our findings suggest publication venues might dedicate time and space to replication studies. This would provide researchers focused on replication efforts with a platform for publication. It is worth noting that some conferences have already taken steps in this direction by including separate submission panels or organizing discussions focused on reproducibility, as observed in \\citet{wilson2012replichi}.", "candidate_comments": ["To address this, we recommend that leaders in publication venues, such as journal editors and conference chairs, to dedicate a section or a panel of their venue replication studies. % research practices  This would provide the researchers focusing on replication efforts with a venue to publish there by increasing the transparency and reproducibility across scientific research. It is worth noting that some conferences started including separate submission panels, or having discussions focused on reproducibility, which is step in right direction \\cite{wilson2012replichi}. "]}
{"annotations": [{"id": 86961817, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86973152, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973703, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_14", "final_parag": "Furthermore, the Indian government has demonstrated its commitment to promoting open science. It is essential that this commitment extends to addressing the replication crisis. Specifically, funding agencies such as the Indian Council of Medical Research (ICMR) and the Department of Science and Technology (DST) should actively encourage the adoption of reproducible research practices.", "candidate_comments": ["Furthermore, the Indian government has shown its commitment to promoting open science. It is essential that they extend the commitment to addressing reproducibility crisis. Specifically, funding agencies such as the Indian Council of Medical Research (ICMR) and the Department of Science and Technology (DST) should actively encourage the adoption of reproducible research practices."]}
{"annotations": [{"id": 87254399, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87259036, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87268260, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_15", "final_parag": "These practices, that prioritize quantity over quality, undermine research quality \\cite{nelson2012let,rahal2023quality}. Specifically, researchers who prioritize transparency and reproducibility may output fewer papers due to time intensive nature of these practices. This could put them at a disadvantage in highly competitive job markets or promotional tracks that prioritize quantity.", "candidate_comments": ["Similarly, promotion decisions are influenced by papers number of publications in high-impact factor journals \\cite{nosek2012scientific, hicks2015bibliometrics}.  These practices, which prioritize quantity over quality, contribute to various research crises by overlooking the importance of research quality\\cite{rahal2023quality}. As noted by Nelson et al.\\cite{nelson2012let}, \"In a world where researchers are rewarded for how many papers they publish, this can lead to a decrease in the truth value of our shared knowledge.\" Researchers who prioritize transparency and reproducibility practices may produce fewer papers due to the time-intensive nature of these practices, potentially putting them at a disadvantage in highly competitive job markets or promotional tracks that prioritize quantity. Therefore, there is a pressing need for a shift in the hiring and promotion processes to place greater emphasis on research transparency \\cite{schonbrodt2016changing}. To facilitate this shift, every university in India should incorporate requirements that integrate candidates' contributions to open science, replicability, and reproducibility into their hiring processes. It is imperative that Indian universities begin to value and recognize professors and scholars who actively promote open and transparent research practices. Such recognition would serve as a meaningful reward for their efforts and encourage others to follow suit."]}
{"annotations": [{"id": 86961572, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86980595, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86980985, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19158-66c30bd111f2a070.jsonl", "id_parag": "2310.19158-66c30bd111f2a070_16", "final_parag": "It is evident that researchers are more includes to trust a tool if its training data, processes, and methodologies for generating the output score are transparent and open. Additionally, researchers stressed the importance of model explainability for building trust and fostering adoption. All participants expressed a preference for hybrid systems that can integrate inputs from both humans and AI, promoting human-AI collaboration.", "candidate_comments": ["Additionally, it is evident that researchers place their trust in a tool when its training data, process and the methodologies for generating the output score are transparent and open. Furthermore, researchers stressed the importance of model explainability of the model and the ability to provide rationale behind the output score as key factors in building trust and adoption.  %The training process of the AI agents needed to be explained properly. There is a lack of trust to accept the estimation provided by the tool because of the lack of explainability. Explainable AI is needed to have more trust for the fully AI version of the tool\\cite{}.  Additionally, all the participants expressed a preference for a system that can integrate inputs from humans as well as AI, leading to human-AI collaboration. % of the tool compared to the fully AI version.  The future work is to develop a human-AI collaboration tool highlighting the contributions of both humans and AI. The tool should remain accessible for all the researchers, allowing them to utilize and validate its functionality. "]}
{"annotations": [{"id": 86734257, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86772636, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86827200, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19159-66c30bd111f2a070.jsonl", "id_parag": "2310.19159-66c30bd111f2a070_1", "final_parag": "The need for a clean and sustainable energy sector has led to significant changes in the modern power grid, including increased integration of renewable energy sources, use of advanced sensors and monitoring devices and growth in electrification~\\cite{IEA_2023}. \nSuch changes have been pivotal in the rise of prosumers (\\ie active energy consumers that produce and consume energy)~\\cite{EEA_prosumer}. \nFor our work, we focus on households and residential prosumers operating with a financial objective of reducing energy bills. \nIn most cases, such residential prosumers rely on PV systems for producing electricity and consuming it instantaneously, with the option of injecting the excess to the power grid.", "candidate_comments": ["We focus on households and residential prosumers with a solar photovoltaic~(PV) system for producing electricity and a home battery for energy storage. These prosumers are driven by the financial objective of reducing energy bills and with increased price volatility, such prosumers are shifting towards more elaborate home energy management systems~(HEMS) As presented in~\\cite{hems_review}, HEMS have been a major field of research, with prior works studying different control strategies for home energy management systems. "]}
{"annotations": [{"id": 87254525, "completed_by": {"id": 70661}, "result": ["Réécriture", "Différent"]}, {"id": 87258768, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent"]}, {"id": 87268059, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}], "agreement": 77.77777777777777, "source_file": "2310.19159-66c30bd111f2a070.jsonl", "id_parag": "2310.19159-66c30bd111f2a070_2", "final_parag": "Driven by increased volatility in the energy markets~\\cite{price-volatility}, we note an increasing shift to storage-based PV systems and more elaborate home energy management systems~(HEMS). \nSupported by increased adoption of sensors, smart meters and other IoT devices, these HEMS make use of advanced control algorithms to identify optimum control strategies for individual prosumers.", "candidate_comments": ["We focus on households and residential prosumers with a solar photovoltaic~(PV) system for producing electricity and a home battery for energy storage. These prosumers are driven by the financial objective of reducing energy bills and with increased price volatility, such prosumers are shifting towards more elaborate home energy management systems~(HEMS) As presented in~\\cite{hems_review}, HEMS have been a major field of research, with prior works studying different control strategies for home energy management systems. ", "MPC is an advanced control strategy that relies on a model of the system and an optimizer~\\cite{mpc_basics}. \nThe optimizer uses the system model to anticipate the future behavior of the system and calculate an optimum control strategy to achieve the required objective."]}
{"annotations": [{"id": 86873352, "completed_by": {"id": 70661}, "result": ["Différent", "Réécriture"]}, {"id": 86901628, "completed_by": {"id": 62306}, "result": ["Réécriture", "Différent"]}, {"id": 86902314, "completed_by": {"id": 62471}, "result": ["Différent", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19159-66c30bd111f2a070.jsonl", "id_parag": "2310.19159-66c30bd111f2a070_3", "final_parag": "This section describes our transfer learning-based forecasting approach and provides details related to the forecasting model (Temporal Fusion Transformer) and the experimental setup used for our simulations.", "candidate_comments": ["We model this electrical demand forecasting problem as a univariate forecasting problem. With the intention of using these forecasts further for control applications, the forecasting problem is modeled as a stochastic forecasting problem.  This leads to a forecasting model~(discussed in~\\cref{sec:methodology}) that provides probabilistic forecasts over the horizon $T$. ", "We present a forecasting approach that utilizes transfer learning and Temporal fusion transformer~(TFT). In this section, we describe these components and detail our experimental setup. "]}
{"annotations": [{"id": 86963067, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86975338, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86975304, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}], "agreement": 55.55555555555556, "source_file": "2310.19159-66c30bd111f2a070.jsonl", "id_parag": "2310.19159-66c30bd111f2a070_4", "final_parag": "This supports our hypothesis that using data from different households in a transfer learning-based approach can lead to data-efficient forecasting models that can produce good quality forecasts and can be used in advanced controllers such as MPCs to develop home energy management systems. \nFollowing up on these results, we plan to expand the scope of our study in two main research areas. The first one focuses on improving the fine-tuning methodology. This involves creating domain-specific finetuning strategies that can leverage the temporal representations learnt by the global model and combine prior knowledge about the household to fine-tune the forecaster more efficiently. The other research direction will focus on open-source contributions. This involves pretraining global models on a larger set of household data and integrating these global models with platforms such as HuggingFace~\\cite{hugging_face}.\\footnote{This is out-of-scope for the work/models presented in this work due to contractual limitations. }", "candidate_comments": ["These results indicate that leveraging data from multiple households to pretrain a global model and then finetuning it for individual households significantly improves demand forecasts. ", "This will enable a wide range of applications and more importantly control architectures, to easily access good quality forecasts and develop home energy managements solutions that can easily scale to new households."]}
{"annotations": [{"id": 86985158, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986046, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86986113, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19168-11fc52765f8eb8fd.jsonl", "id_parag": "2310.19168-11fc52765f8eb8fd_1", "final_parag": "The results of our reconstructions are not fully imperative for species classification and mapping, since our models are trained with a contrastive objective. In the wild, animals come in different poses and sizes, making it challenging to reconstruct them perfectly. However, our models have effectively learned the structures of different bird species. The large scale pre-training along-with satellite imagery and metadata such as month of year and location helps our model learn robust fine-grained features. The zero-shot reconstruction results on CUB-200-2011 and NABirds (Figure \\ref{fig:cubrecon} and Figure \\ref{fig:narecon}) confirm that the features learned by our models are highly transferable.", "candidate_comments": ["Since our models are trained with a contrastive objective, having \\texttt{perfect} reconstructions is not fully important for species classification and mapping. Reconstructing animals in the wild is hard since they vary in pose and size. Yet, our models are able to learn the structure of various bird species effectively. The zero-shot reconstruction results~(Figure \\ref{fig:cubrecon} and Figure \\ref{fig:narecon}) on CUB-200-2011 and NABirds reaffirm that the features learned by our models are highly transferable."]}
{"annotations": [{"id": 86738330, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86901905, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86902485, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19168-e1425d772d0e14e8.jsonl", "id_parag": "2310.19168-e1425d772d0e14e8_1", "final_parag": "We propose a metadata-aware self-supervised learning~(SSL)~framework useful for fine-grained classification and ecological mapping of bird species around the world. Our framework unifies two SSL strategies: Contrastive Learning~(CL) and Masked Image Modeling~(MIM), while also enriching the embedding space with metadata available with ground-level imagery of birds. We separately train uni-modal and cross-modal ViT on a novel cross-view global bird species dataset containing ground-level imagery, metadata (location, time), and corresponding satellite imagery. We demonstrate that our models learn fine-grained and geographically conditioned features of birds, by evaluating on two downstream tasks: fine-grained visual classification~(FGVC) and cross-modal retrieval. Pre-trained models learned using our framework achieve SotA performance on FGVC of iNAT-2021 birds and in transfer learning settings for CUB-200-2011 and NABirds datasets. Moreover, the impressive cross-modal retrieval performance of our model enables the creation of species distribution maps across any geographic region. The dataset and source code will be released at~\\url{https://github.com/mvrl/BirdSAT}. \n\\end{abstract}", "candidate_comments": ["Large scale self-supervised learning (SSL) methods have been proven to generalize well over a variety of downstream tasks.  Keeping this in mind, we aim to create a unified transformer-based model which is capable of classifying and mapping bird species across the globe. \nWe consider the problem of fine-grained visual classification of birds around the world for which habitat is an essential contextual information. To leverage this information, we construct a large scale cross-view dataset consisting of paired ground level images of birds and corresponding satellite images. The additional visual modality not only aids in visual recognition of birds but also enables the construction of global maps of bird distributions. We evaluate uni-modal and cross-modal training of vision transformers on this dataset. More specifically, we incorporate metadata along with contrastive and masked reconstruction objectives. We explore the ability of these models in encoding fine-grained concepts by demonstrating their effectiveness on fine-grained classification and zero-shot retrieval tasks. We find that our proposed method outperforms the state-of-the-art method in fine-grained classification of iNAT-2021 birds. Furthermore, it improves over the baselines on CUB-200-2011 and NABirds datasets. The dataset and source code are released at \\url{https://github.com/TBD}."]}
{"annotations": [{"id": 86972473, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86985122, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985268, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19168-e1425d772d0e14e8.jsonl", "id_parag": "2310.19168-e1425d772d0e14e8_2", "final_parag": "Species classification and distribution mapping are two important tasks for ecologists who monitor and protect the habitats of endangered species. Species classification involves categorizing species with subtle differences into fine-grained classes. It lies within the continuous manifold of basic visual recognition tasks and more complex visual identification tasks. Moreover, it coincides with the task of Fine-Grained Visual Classification (FGVC) which has already been used for distinguishing between models of cars~\\cite{Akrause20133d}, species of birds~\\cite{van2018inaturalist,van2015building,wah2011caltech}, airplanes~\\cite{maji2013fine}, etc. On the other hand, the task of species distribution mapping aims at mapping the habitation of species of interest over any geographic region in the world. In this work, we propose to learn a unified representation space useful for solving both of these tasks. Specifically, we evaluate our framework for classifying and mapping \\texttt{bird} species around the world. However, our models are general enough to be easily extended to~\\texttt{any} species of interest.", "candidate_comments": ["Srikumar: Species classification %Fine-Grained Visual Classification (FGVC) is a long-standing research problem in computer vision and is an important task for ecologists who monitor and protect habitats of endangered species. It involves categorizing species with subtle differences into fine-grained classes. More generally, it lies within the continuous manifold of basic visual recognition tasks and more complex visual identification tasks. Moreover, it coincides with the task of Fine-Grained Visual Classification (FGVC) which has already been used for distinguishing between models of cars~\\cite{Akrause20133d}, species of birds~\\cite{van2018inaturalist,van2015building,wah2011caltech} and airplanes~\\cite{maji2013fine}.Lately, species classification is used in several commercial technologies and applications.%"]}
{"annotations": [{"id": 86828551, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86856684, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86898714, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19168-e1425d772d0e14e8.jsonl", "id_parag": "2310.19168-e1425d772d0e14e8_3", "final_parag": "\\begin{itemize} \\item We introduce a global Cross-View iNAT 2021 Birds Dataset, which contains paired satellite images and corresponding ground-level bird images. \n\\item We propose a framework for cross-view pre-training of vision transformers along with metadata enabling the ecological mapping of bird species. \n\\item We demonstrate the rich representational capability of our pre-trained models by demonstrating SotA on fine-grained bird classification across three datasets. \n\\end{itemize}", "candidate_comments": ["In this study, we focus on fine-grained visual classification of iNAT-2021 Birds, for which rich metadata is made available.  We leverage the power of self-supervised pre-training using Masked Autoencoders (MAE) to learn rich visual features. Yet, "]}
{"annotations": [{"id": 86735078, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86900254, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86900367, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19181-4dae017aa913c5a4.jsonl", "id_parag": "2310.19181-4dae017aa913c5a4_1", "final_parag": "We generated heatmaps to visualize human annotators' evaluations of prompts generated by GPT-3.5T in Figure~\\ref{fig:heatmap_groundtruth_gpt35} and GPT-4 in Figure~\\ref{fig:heatmap_groundtruth_gpt4}.", "candidate_comments": ["The heatmap for GPT-3.5-Turbo generated prompts is presented in Figure~\\ref{fig:heatmap_groundtruth_gpt35} and the heatmap for GPT-4 generated prompts is presented in Figure~\\ref{fig:heatmap_groundtruth_gpt4}."]}
{"annotations": [{"id": 87254668, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87259203, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87268314, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19181-4dae017aa913c5a4.jsonl", "id_parag": "2310.19181-4dae017aa913c5a4_2", "final_parag": "We selected both malicious and benign prompt collections, ensuring each individual prompt is labeled, from our common groundtruth dataset. Upon merging the two collections, we observed a total of 1,255 malicious prompts and 3,123 benign prompts. Recognizing this imbalance in the distribution of prompts, we opted to exclude some benign prompt collections. In total, we considered 50 benign prompt collections along with 258 malicious prompt collections. We split this dataset into 70\\% for training, 20\\% for testing, and 10\\% for validation, by maintaining the balance mentioned above.", "candidate_comments": ["Table~\\ref{tab:groundtruth Individualprompts} shows the summary of groundtruth for \\emph{individual prompt detection}.  This groundtruth is not completely but almost balanced. This is because the number of prompts in collections varies and with random selection of collections, we do not obtain the same number of prompts.  \\shirin{the groundtruth is not completely balanced, so we should acknowledge and justify it.}  Analyzing the groundtruth for \\emph{individual prompt detection}, we observed that it is not completely balanced. The imbalance in dataset reflects its complex nature of real-world data, where malicious prompts are naturally less frequent than benign ones. This aspect of dataset allows our model to train in an environment such as real-time scenarios. "]}
{"annotations": [{"id": 86870525, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86904045, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86910138, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19181-4dae017aa913c5a4.jsonl", "id_parag": "2310.19181-4dae017aa913c5a4_3", "final_parag": "To test the capability of our individual classifier for detecting malicious collections, we created a balanced test set with 50 malicious and 53 benign collections.", "candidate_comments": ["This includes 50 malicious prompt collections and 53 benign prompt collections.  Along with the existing attributes like Collection Number, Prompt Number, Prompt, Prompt Label and Version, we used an additional attribute named as \"Collection Label\". "]}
{"annotations": [{"id": 86971957, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86985938, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86988991, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19181-4dae017aa913c5a4.jsonl", "id_parag": "2310.19181-4dae017aa913c5a4_4", "final_parag": "To evaluate the individual prompt classifier performance on entire collection of prompts, we presented collections of prompts during testing. To obtain collections, we concatenate all the prompts within a collection, based on their prompt numbers, to form a single paragraph. This consolidated input is then pre-processed and presented to the model to obtain a prediction.", "candidate_comments": ["We use this attribute for evaluating the models performance on entire collection of prompts.  Finally we split this curated dataset into 70\\% training, 20\\% testing and 10\\% for validation and proceed with the model finetuning step.  Details about this distribution are provided in Table~\\ref{tab:groundtruth wholecollection}. "]}
{"annotations": [{"id": 86829213, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86856575, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86898666, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19181-4dae017aa913c5a4.jsonl", "id_parag": "2310.19181-4dae017aa913c5a4_5", "final_parag": "\\textbf{Challenges with Whole Collection Classification:} Based on the results obtained after evaluating model's performance with entire collection, we can observe that conversations provide more nuanced understanding of users intent compared to evaluating individual prompts separately. This comprehensive approach increases the chance of detecting malicious activity. However, obtaining and testing an entire prompt collection is not feasible in real-time, where users interact with chatbots one prompt at a time. To adapt to real-time scenarios, we propose to examine the current prompt alongside with its preceeding prompts, to ascertain if the context captured unveils any malicious activity.", "candidate_comments": ["\\textbf{Comparison with Detection Scheme-1:}  A comparative analysis between scheme-2, which involves whole collection detection, and scheme-1, focusing on individual prompt detection, showcases the efficacy of scheme-2. Based on the results obtained after evaluating model's performance with entire collection, we can observe that conversations provide more nuanced understanding of users intent compared to evaluating individual prompts separately. This comprehensive approach increases the chance of detecting malicious activity."]}
{"annotations": [{"id": 86960126, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86965004, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86965109, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19181-50b3a1199222f948.jsonl", "id_parag": "2310.19181-50b3a1199222f948_1", "final_parag": "The motivation behind implementing these attacks is to cover a diverse range of phishing websites that have been detected and studied in the literature. By investigating the capability of the LLMs to generate these attacks, we aim to demonstrate its potential impact on the security landscape and raise awareness among security researchers and practitioners.", "candidate_comments": ["\\shirin{ideally, we should explain these attacks, or at least name them. } By investigating the capability of ChatGPT to generate these types of attacks, we aim to demonstrate its potential impact on the security landscape and raise awareness among security researchers and practitioners. "]}
{"annotations": [{"id": 86961128, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86975175, "completed_by": {"id": 126844}, "result": ["Différent", "Réécriture"]}, {"id": 86975243, "completed_by": {"id": 62306}, "result": ["Réécriture", "Différent"]}], "agreement": 77.77777777777777, "source_file": "2310.19181-50b3a1199222f948.jsonl", "id_parag": "2310.19181-50b3a1199222f948_2", "final_parag": "We examined the number of iterative prompts required by three independent coders (two graduate students and one undergraduate student in Computer Science) to create each of the phishing attacks described in Table~\\ref{table:summary_phishing_attacks}. The coders possessed varying levels of technical proficiency in Computer Security: Coder 1 specialized in the field, Coder 2 had good experience, and Coder 3 had some familiarity through academic coursework. Table~\\ref{tab:prompts_required} presents the average number of prompts required across the three coders to generate the phishing functionality (attacks) across all four LLM models. \nEach coder created their own set of prompts for designing the website layout and for transmitting the stolen credentials back to the attacker, which they reused for multiple attacks.", "candidate_comments": ["Finally, the LLM can be asked to create a function to send the credentials entered on the phishing websites to the attacker. LLMs can generate essential JS functions or PHP scripts to send the credentials using email, send it to an attacker-owned remote server, or store it in a back-end database. ", "\\shirin{add a conclusion from the table. What is the outcome/ goal of this analysis? What did we see? } \\shirin{the table can be confusing. Needs some explanation on the use of +. Is it additional prompts? should we add the total average at the end?}  \\shirin{I see. you will be explaining it later.} for designing the website layout and for transmitting the stolen credentials back to the attacker."]}
{"annotations": [{"id": 86899314, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86977778, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86978831, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}], "agreement": 55.55555555555556, "source_file": "2310.19181-591e27c149e249d7.jsonl", "id_parag": "2310.19181-591e27c149e249d7_1", "final_parag": "\\subsection{Effectiveness of generated content} We explored the proficiency of commercial LLMs in generating both phishing websites and emails. To assess phishing websites, we began with a brief case study on the effort necessary to craft prompts manually. These prompts are designed to guide each of the four commercial LLMs in producing functional phishing websites with their respective attack vectors. While manual prompt generation is insightful, the potential for scalable attacks hinges on automatically created prompts. Thus, we conducted a qualitative evaluation of the quality of websites produced by such automated prompts. To further gauge the efficacy of these LLM-generated attacks, we contrasted the reactions of popular anti-phishing blocklists to traditional phishing attacks and those generated by LLMs, focusing on coverage and detection speed. For assessing phishing emails, we employed four text generation metrics: BLEU, Rouge, Topic Coherence, and Perplexity. Using these metrics, we compared the email text generated by each commercial LLM model to the original human-crafted versions. \n\\subsection{Automated detection of phishing prompts} After assessing the potential exploitation of commercial LLMs in generating phishing scams at scale, we designed a machine learning-based detection model to prevent LLMs from producing such malicious content. To build our ground truth, we manually labeled prompts associated with phishing website generation. To explore the best detection method, we tested our finetuned model using three different approaches: individual prompt detection, entire collection detection, and prompt subsets detection. In all these approaches, we finetuned a pre-trained RoBERTa~\\cite{liu2019roberta} using a groundtruth dataset with individual prompts and tested its capability across individual prompts, entire collections, and prompt subsets.", "candidate_comments": ["\\subsection{Evaluating generated attacks:} In Section~\\sayak{add}, we qualitatively evaluate the content generated by both manually-designed and model-manufactured prompts at generating phishing content, with the aim of assessing their respective capacities in crafting phishing websites and emails.  Furthermore, we engage in a comparative study of the variations in performance among different LLMs and configurations, pinpointing specific attributes or features that are present in the generated phishing content. To evaluate the quality of the generated phishing emails, we check their appearance and functionality. Additionally, we also utilized four text generation metrics such as BLEU, Rogue, Topic Coherence, and Perplexity to quantitatively measure the performance of the LLM-generated text with the ground-truth dataset.", "\\subsection{Automated detection of malicious prompts:} In Section~\\sayak{add} we utilize our findings to develop a machine learning-aided model that can detect malicious prompts that can be used to generate phishing websites or emails. Our model specializes in detecting single prompts, as detailed in Section~\\sayak{cite} as well as early detection of multi-stage prompts to prevent attackers from generating malicious content as early as possible~\\sayak{cite}. \n\\sayak{@Poojitha, extend}"]}
{"annotations": [{"id": 86904344, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86971760, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86972257, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19181-6f447c12908e62c2.jsonl", "id_parag": "2310.19181-6f447c12908e62c2_1", "final_parag": "Recent works treat phishing email and spam detection as a text classification task and utilize pre-trained language models, such as BERT~\\cite{devlin2018bert}, to detect phishing emails~\\cite{otieno2023application, karki2022using} and spam~\\cite{rifat2022bert, oswald2022spotspam}.", "candidate_comments": ["Not limited to the task of detecting phishing emails, pre-trained language models like BERT have been used for detecting spams~\\cite{rifat2022bert}\\cite{oswald2022spotspam} treating a spam as text involving malicious intent. "]}
{"annotations": [{"id": 86962894, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86979747, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86979922, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19181-6f447c12908e62c2.jsonl", "id_parag": "2310.19181-6f447c12908e62c2_2", "final_parag": "A couple of works have also utilized pre-trained language models for detecting phishing websites from the URLs~\\cite{he2023method,wang2023large}.", "candidate_comments": ["Not limited to the task of detecting phishing emails, pre-trained language models like BERT have been used for detecting spams~\\cite{rifat2022bert}\\cite{oswald2022spotspam} treating a spam as text involving malicious intent. "]}
{"annotations": [{"id": 86733111, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86903849, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86906565, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19181-6f447c12908e62c2.jsonl", "id_parag": "2310.19181-6f447c12908e62c2_3", "final_parag": "However, our approach focuses on a more preventive strategy. Instead of concentrating on detecting malicious content after its generation, our main objective is to obstruct the generation of harmful codes by the LLMs. We aim to examine and filter the prompts by hindering the creation of malicious content before it starts.", "candidate_comments": ["These works show a comparative analysis with traditional algorithms on SMS Spam dataset and proved that pretrained language models excel in understanding the malicious intent.  "]}
{"annotations": [{"id": 86982099, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86982813, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86984082, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19181-db20a9efc6b1a3e2.jsonl", "id_parag": "2310.19181-db20a9efc6b1a3e2_1", "final_parag": "\\section{Conclusion} \\subsection{Implications} Our research indicates that readily available commercial LLMs can effectively generate phishing websites and emails. These LLMs are not only capable of being manually directed to initiate attacks but can also autonomously generate phishing prompts. These AI-created prompts are adept at producing phishing content that can evade current anti-phishing tools as effectively as human-generated content. Moreover, phishing emails derived from LLM-generated prompts can mimic authentic human phishing attempts with high accuracy. The potential misuse of LLMs for phishing poses a serious threat, as attackers can refine and reuse a small set of prompts to create a vast number of sophisticated phishing attacks. However, we have developed a machine learning model that can detect these malicious prompts early on, which is crucial in preempting the production of harmful content by LLMs. Our model, which demonstrates strong performance in identifying phishing prompts for both websites and emails, could be integrated with LLMs as a third-party plugin.", "candidate_comments": ["Our research reveals that widely accessible commercial LLMs can proficiently produce phishing websites and emails. While these models can be manually prompted to launch attacks, we have discovered a more sophisticated method: using LLMs to autonomously craft prompts for phishing scams. Furthermore, not only do LLM-crafted prompts excel in creating phishing content, but the resulting websites prove as evasive to anti-phishing tools as those manually designed by humans. Similarly, phishing emails generated through this method can convincingly emulate the style and content of human-composed phishing emails. This misuse of commercial LLMs presents severe potential dangers. Attackers can easily iterate over a handful of optimized prompts, enabling them to generate a limitless array of malicious prompts to exponentially amplify their attacks. An effective countermeasure seems to be the early detection of these malevolent prompts, preventing the LLM from producing harmful content. In light of this, we developed a machine learning model that performs well in identifying both phishing websites and email prompts. Our detection approach can potentially prevent the LLM from generating malicious content. Our model can potentially be integrated with LLMs as a third-party plugin, necessarily preventing attackers from utilizing commercial LLMs as a source for generating phishing scams. Additionally, the dataset used for training our machine learning model provides a novel source of annotated phishing prompts that can further drive research in this space. "]}
{"annotations": [{"id": 86731097, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86726679, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86728387, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19188-ae2a90d4f1c778ed.jsonl", "id_parag": "2310.19188-ae2a90d4f1c778ed_1", "final_parag": "We refer to our method as unsupervised, meaning that it does not require 3D data to perform 3D reconstruction. Thus, using other features/outputs from models trained without 3D data do not alter the unsupervised 3D reconstruction setup. \nWe do, however, restrain from using keypoints/pose estimators as they limit our approach to category-specific conventions.", "candidate_comments": ["A reflection on the 2D and 3D performance gaps provokes a compelling question: Given a large-scale image dataset with no preliminary knowledge, can we utilise much of the maturing techniques in 2D to extract features beneficial for truly unsupervised and generalisable 3D reconstruction? To this end, we propose an end-to-end pipeline requiring only a self-supervised DINO-ViT model pretrained on ImageNet to reconstruct implicit 3D representations on novel categories."]}
{"annotations": [{"id": 86726742, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86727491, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86728750, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19188-ae2a90d4f1c778ed.jsonl", "id_parag": "2310.19188-ae2a90d4f1c778ed_2", "final_parag": "In summary, our contributions are: \\emph{(i)} a new problem formulation on mining 3D shapes from large-scale web-retrieved images without any priors or annotations; \\emph{(ii)} an end-to-end pipeline to cluster, estimate pose, and generate neural 3D representations from unannotated image datasets without any 3D ground truths; \\emph{(iii)} a detailed empirical study to showcase the superiority of our method on challenging datasets and robustness under categories where no previous ground-truth reconstructions exist.", "candidate_comments": ["\\begin{itemize}     \\item A new problem formulation on mining 3D from large-scale web-retrieved images without any priors or annotations. \n    \\item An end-to-end pipeline to cluster, estimate pose, and deduce accurate 3D representations from unannotated image datasets via only 2D self-supervision. \n    \\item A detailed empirical study on our approach to showcase the superiority of our method in challenging datasets and robustness under categories where no previous ground-truth reconstructions exist. \n\\end{itemize}"]}
{"annotations": [{"id": 86983082, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86985248, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86985237, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19199-2c7d26c27da0bcad.jsonl", "id_parag": "2310.19199-2c7d26c27da0bcad_1", "final_parag": "\\begin{abstract} We propose a 3D simulator tailored for the Drone-as-a-Service framework. The simulator enables employing dynamic algorithms for addressing realistic delivery scenarios. We present the simulator's architectural design and its use of an energy consumption model for drone deliveries. We introduce two primary operational modes within the simulator: the edit mode and the runtime mode. Beyond its simulation capabilities, our simulator serves as a valuable data collection resource, facilitating the creation of datasets through simulated scenarios. Our simulator empowers researchers by providing an intuitive platform to visualize and interact with delivery environments. Moreover, it enables rigorous algorithm testing in a safe simulation setting, thus obviating the need for real-world drone deployments. Demo: \\url{https://youtu.be/HOLfo1JiFJ0}", "candidate_comments": ["original We propose a 3D simulation tool tailored for the Drone-as-a-Service framework. We present the tool's architectural design and its utilization of an energy consumption model for drone deliveries. The tool employs real-time dynamic algorithms for addressing realistic delivery scenarios. We introduce two primary operational modes within the tool: the edit mode and the runtime mode. Beyond its simulation capabilities, our tool serves as a valuable data collection resource, facilitating the creation of datasets through simulated scenarios. Its primary purpose is to empower researchers by providing an intuitive platform to visualize and interact with delivery environments, enabling rigorous algorithm testing in a secure simulation setting, thus obviating the need for real-world drone deployments. Demo: \\url{https://youtu.be/0H-soTNdTBk}"]}
{"annotations": [{"id": 86723864, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86731546, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86729658, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19199-2c7d26c27da0bcad.jsonl", "id_parag": "2310.19199-2c7d26c27da0bcad_2", "final_parag": "Furthermore, we introduce the concept of \\textit{waypoints}, as depicted in Fig. \\ref{fig:editMode}. Waypoints connect segments that cannot be directly linked to two nodes due to disparities in elevation or challenging terrain.", "candidate_comments": ["original Furthermore, we introduce the concept of waypoints, as depicted in Fig. \\ref{fig:editMode}, which serves to connect segments that cannot be directly linked to two nodes due to disparities in elevation or challenging terrain. This enhancement expands upon the concept initially presented in \\cite{alkouz2020swarm}. It adds a layer of realism by acknowledging that drones cannot always establish a direct, unobstructed flight path between nodes. If the settings button is pressed, users can specify attributes related to the energy consumption model, payload capacity, and drone speed (Fig. \\ref{fig:settings}). "]}
{"annotations": [{"id": 86965033, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986610, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86986551, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19199-2c7d26c27da0bcad.jsonl", "id_parag": "2310.19199-2c7d26c27da0bcad_3", "final_parag": "This concept adds a layer of realism by acknowledging that drones cannot always establish a direct, unobstructed flight path between nodes \\cite{alkouz2022flight}. Lastly, when the settings button is pressed, users can specify attributes related to the energy consumption model, payload capacity, and drone speed (Fig. \\ref{fig:settings}).", "candidate_comments": ["original Furthermore, we introduce the concept of waypoints, as depicted in Fig. \\ref{fig:editMode}, which serves to connect segments that cannot be directly linked to two nodes due to disparities in elevation or challenging terrain. This enhancement expands upon the concept initially presented in \\cite{alkouz2020swarm}. It adds a layer of realism by acknowledging that drones cannot always establish a direct, unobstructed flight path between nodes. If the settings button is pressed, users can specify attributes related to the energy consumption model, payload capacity, and drone speed (Fig. \\ref{fig:settings}). "]}
{"annotations": [{"id": 86982383, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985778, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86985721, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19199-2c7d26c27da0bcad.jsonl", "id_parag": "2310.19199-2c7d26c27da0bcad_4", "final_parag": "\\subsubsection{Runtime Mode:} Once the play button is hit in the Edit mode, the Runtime mode starts (Fig. \\ref{fig:runtime}). In the runtime mode, the drones become visible and begin traversing according to the provided algorithm. As previously mentioned, this tool operates in a non-deterministic manner. Consequently, if a segment (a service) becomes inaccessible from the originally composed path, the algorithm promptly adapts to this change, redirecting the drones accordingly.\\looseness=-1", "candidate_comments": ["original Once the play button is hit in the Edit mode, the Runtime mode starts (Fig. \\ref{fig:runtime}). In the runtime mode, the drone models become visible, and they begin traversing according to the provided algorithm. As previously mentioned, it is essential to note that this tool operates in a non-deterministic manner. Consequently, if a segment (a service) becomes inaccessible from the originally composed path, the algorithm promptly adapts to this environmental change, redirecting the drones accordingly."]}
{"annotations": [{"id": 86900929, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86901775, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86906768, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19199-2c7d26c27da0bcad.jsonl", "id_parag": "2310.19199-2c7d26c27da0bcad_5", "final_parag": "The camera tracks the drones during their journey automatically. However, users also have the option to take control using the WASD buttons for manual camera adjustments. To provide users with an enriched experience, double-clicking on any item in the environment offers a close-up view. Each object within the simulation is thoughtfully labeled to convey its relevant attributes, such as segment length or drone battery percentage. The timer displays the simulation time elapsed since the commencement of the trip and permits users to adjust the simulation speed, enabling the drones to move at varying rates.", "candidate_comments": ["original The camera tracks the drones during their journey, but users also have the option to take control using the WASD buttons for manual camera adjustments. To provide users with an enriched experience, double-clicking on any item in the environment offers a close-up view. Each object within the simulation is thoughtfully labeled to convey its relevant attributes, such as segment length or drone battery percentage. The timer prominently displays the simulation time elapsed since the commencement of the trip and permits users to adjust the simulation speed, enabling the drones to move at varying rates."]}
{"annotations": [{"id": 86737180, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86837213, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86898536, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19199-2c7d26c27da0bcad.jsonl", "id_parag": "2310.19199-2c7d26c27da0bcad_6", "final_parag": "This tool serves a dual purpose: not only it assists users in visualizing potential scenarios in drone deliveries under the DaaS model but also functions as a valuable data collection tool for environmental interactions. Following each trip, users may export a CSV file containing comprehensive data regarding the status of every drone in the network. This data encompasses crucial information such as drone energy consumption, battery levels, time spent at each node, and travel times for each segment at each time frame. Such data can be leveraged to refine and improve DaaS composition algorithms.", "candidate_comments": ["csv As previously mentioned, the simulator enables data collection for environmental interactions. Following each trip, users can export a CSV file containing comprehensive data regarding the status of every drone in the network. This data encompasses crucial information such as drone energy consumption, battery levels, time spent at each node, and travel times for each segment at each time frame. Such data can be leveraged to refine and improve DaaS composition algorithms."]}
{"annotations": [{"id": 86962593, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86962685, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86963654, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19204-784ace62db0ce395.jsonl", "id_parag": "2310.19204-784ace62db0ce395_1", "final_parag": "\\begin{abstract} While ChatGPT is a well-known artificial intelligence chatbot being used to answer human's questions, one may want to discover its potential in advancing software testing. We examine the capability of ChatGPT in advancing the intelligence of software testing {through a case study} on metamorphic testing {(MT)}, a state-of-the-art software testing {technique}. We {ask} ChatGPT to generate {candidates of} metamorphic relations (MRs), which are basically necessary properties of the {object program and which traditionally require} human intelligence to identify. These {MR candidates} are then evaluated in terms of correctness by domain experts. \nWe show {that ChatGPT} can be used to generate new {correct MRs} to test several software systems. Having said that, the majority of {MR candidates} are either defined vaguely or incorrect, especially for {systems that {have} never been tested with {MT}}. ChatGPT can be used to advance software testing intelligence by {proposing} MR candidates that can be later adopted for implementing tests; but human intelligence should still {inevitably} be involved to justify and {rectify their} correctness.", "candidate_comments": ["Generative Pre-trained Transformer Chatbot, namely ChatGPT, is a new intelligent agent launched by OpenAI in November 2022. It immediately recognized worldwide for its extraordinary intelligence which is praised by the New York Times ``the best artificial intelligence chatbot ever released to the general public''. In this paper, we examine the performance of ChatGPT in generating metamorphic relations (MRs), a challenge task that is traditionally require human knowledge in the identification. "]}
{"annotations": [{"id": 86899080, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86900342, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86907889, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19210-66c30bd111f2a070.jsonl", "id_parag": "2310.19210-66c30bd111f2a070_1", "final_parag": "\\begin{itemize} \\item We employ weak and strong data augmentations to generate two sufficiently different views. \n\\item We introduce a co-training-based framework to effectively learn embeddings of unlabeled data. \n\\item We propose a community detection approach to clustering assignments that accurately detects novel categories of unlabeled instances while simultaneously determining the number of clusters. \n\\item Through experimentation on six benchmarks, our approach demonstrates state-of-the-art performance and confirms the efficacy of our method. \n\\end{itemize}", "candidate_comments": ["More directly, instances belonging to the same semantic category  should be drawn closer together in the feature space, instead of indiscriminately pushing away the embeddings of all unlabeled samples. "]}
{"annotations": [{"id": 86828156, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86856129, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86898482, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19210-66c30bd111f2a070.jsonl", "id_parag": "2310.19210-66c30bd111f2a070_2", "final_parag": "\\textbf{Novel Category Discovery.} The objective of Novel Category Discovery (NCD) is to cluster unlabeled instances using transferable knowledge derived from the labeled dataset of known categories, with all unlabeled instances regarded as novel categories \\cite{DBLP:journals/corr/abs-2302-12028}. \nKCL \\cite{DBLP:conf/iclr/HsuLK18} addresses the problem of transfer learning across domains, especially cross-task transfer learning, which is similar with NCD. Similar to KCL, MCL \\cite{DBLP:conf/iclr/HsuLSOK19} constructs pairwise pseudo-labels on unlabeled dataset using a similarity prediction network trained on the labeled dataset, and determines whether two instances belong to the same category by using the inner product. DTC \\cite{DBLP:conf/iccv/HanVZ19} utilizes a cross-entropy classifier to initialize representations on the labeled dataset. Then it maintains a list of class prototypes that represent cluster centers, and assigns instances to the closest prototype. In this way, a good clustering representation is learned. \nRankStats or AutoNovel \\cite{DBLP:conf/iclr/HanREVZ20,DBLP:journals/pami/HanREVZ22} applies RotNet \\cite{DBLP:conf/iclr/GidarisSK18} to perform an initialization of the encoder. Afterward, it generates pseudo-labels through dual-ranking statistics and enforces consistency between the predictions of the two branch networks. NCL \\cite{DBLP:conf/cvpr/ZhongFRL0S21} retrieves and aggregates pseudo positive pairs by contrastive learning, and produces hard negative pairs by mixing labeled and unlabeled instances in the feature space. Unlike the methods that separately design loss functions for labeled and unlabeled samples, UNO \\cite{DBLP:conf/iccv/FiniSLZN021} introduces a unified objective function to discover new classes that effectively supports the collaboration between supervised and unsupervised learning. MEDI \\cite{DBLP:conf/iclr/ChiLYLL00ZS22} proposes a new approach using MAML \\cite{DBLP:conf/icml/FinnAL17} to solve the NCD problem and provided a solid theoretical analysis of the underlying assumptions in the NCD field. ComEx \\cite{DBLP:conf/cvpr/YangZYWD22} modifies UNO \\cite{DBLP:conf/iccv/FiniSLZN021} by designing two groups of experts to learn the entire dataset in a complementary way, thus mitigating the limitations of previous NCD methods. Recently, GCD \\cite{DBLP:conf/cvpr/VazeHVZ22} extends the NCD task to include the recognition of known classes in the unlabeled dataset. It leverages a well-trained visual transformer model \\cite{DBLP:conf/iclr/DosovitskiyB0WZ21} to improve visual representation. \nORCA \\cite{DBLP:conf/iclr/CaoBL22} enables the simultaneous clustering of various novel classes in the unlabeled set, assuming a pre-defined number of novel classes.", "candidate_comments": ["GPC \\cite{DBLP:journals/corr/abs-2305-06144} proposes a variant of semi-supervised Gaussian mixture model that dynamically determines prototypes by checking the compactness and separability of clustering, and learns prototype contrastive loss under labeled constraints."]}
{"annotations": [{"id": 86985770, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86987355, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86989893, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-66c30bd111f2a070.jsonl", "id_parag": "2310.19220-66c30bd111f2a070_1", "final_parag": "\\begin{abstract} The dynamic pricing problem has been extensively studied under the {\\bf stream} model: A stream of customers arrives sequentially, each with an independently and identically distributed valuation. \nHowever, this formulation is not entirely reflective of the real world. In many scenarios, high-valuation customers tend to make purchases earlier and leave the market, leading to a {\\em shift} in the valuation distribution. \nThus motivated, we consider a model where a {\\bf pool} of $n$ non-strategic unit-demand customers interact repeatedly with the seller. \nEach customer monitors the price intermittently according to an independent Poisson process and makes a purchase if the observed price is lower than her {\\em private} valuation, whereupon she leaves the market permanently.", "candidate_comments": ["with {\\em optimal} worst-case performance guarantees. \ncompetitive ratio, in terms of both the number of prices and aspect ratio (i.e., ratio between the highest and lowest prices)."]}
{"annotations": [{"id": 86904801, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972989, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973388, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19220-66c30bd111f2a070.jsonl", "id_parag": "2310.19220-66c30bd111f2a070_2", "final_parag": "We present an efficient algorithm that computes a non-adaptive policy whose expected revenue is at least $1/k$ times the optimum, where $k$ is the number of prices. This result is {\\em optimal}: We show that there exists a family of instances on $k$ prices such that no algorithm guarantees more than $1/k$ fraction of optimal revenue. \nFurthermore, the output of our algorithm also guarantees a $1/\\log \\rho$ fraction of the optimum where $\\rho$ is the aspect ratio, i.e., the ratio of the highest and lowest prices. \nFinally, we propose a learn-then-earn type policy with $\\tilde O(n^{3/4})$ regret, i.e., the loss due to not knowing the true valuations.\\fi \\end{abstract}", "candidate_comments": ["with {\\em optimal} worst-case performance guarantees. \ncompetitive ratio, in terms of both the number of prices and aspect ratio (i.e., ratio between the highest and lowest prices)."]}
{"annotations": [{"id": 86872856, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86901057, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86901126, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-79afd71960e8496f.jsonl", "id_parag": "2310.19220-79afd71960e8496f_1", "final_parag": "Most existing work on dynamic pricing employs what we call a {\\bf stream} model: A stream of customers arrives sequentially, each with an independent identically distributed (i.i.d.) valuation. Demand uncertainty is well understood under this model; see, e.g., \\cite{kleinberg2003value,besbes2009dynamic} and \\cite{babaioff2015dynamic}.", "candidate_comments": ["This problem falls under the umbrella of dynamic pricing with unknown demand, which has been extensively studied, see e.g., surveys by \\cite{elmaghraby2003dynamic,den2015dynamic}."]}
{"annotations": [{"id": 86963204, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86972792, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974959, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19220-79afd71960e8496f.jsonl", "id_parag": "2310.19220-79afd71960e8496f_2", "final_parag": "Consider the sale of new clothing lines. \nEach customer visits the (online or offline) store intermittently depending on their availability and makes a purchase if the observed price is lower than her valuation. As each customer typically needs only one unit, she exits the market once a purchase is made. As the product is newly introduced, the seller has little information about customers' valuations to inform their pricing strategy upfront.", "candidate_comments": ["Each independently observes the price and makes her purchase decision if the price is lower than her private valuation. \nWe refer readers to the survey conducted by \\cite{den2015dynamic} for an overview."]}
{"annotations": [{"id": 86734872, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86899477, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86900137, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-79afd71960e8496f.jsonl", "id_parag": "2310.19220-79afd71960e8496f_3", "final_parag": "Orthogonal to non-stationarity, the {\\em independence} assumption is also questionable. In the stream model, the demand in every time period is independent of the previous prices (even if the demand function is non-stationary over time). \nHowever, this is not true in the previous example. \nTo see why the demand rate depends on previous prices, \\sps\\ all customers have a valuation $0.5$ and {\\em always} monitor the price. \nThen, the demand rate is $0$ at price $1$ if and only if the price has {\\em ever} been lower than $0.5$.", "candidate_comments": ["For example, if the price is always zero, then all customers would quickly buy and leave the market."]}
{"annotations": [{"id": 86964463, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86973147, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973692, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19220-79afd71960e8496f.jsonl", "id_parag": "2310.19220-79afd71960e8496f_4", "final_parag": "In order to address these challenges, we consider a single-item revenue maximization problem where a {\\bf pool} of unit-demand, non-strategic customers interact repeatedly with a single seller. \nEach customer monitors the price intermittently according to an independent Poisson process and makes a purchase if she observes a price lower than her private valuation, whereupon she leaves the market permanently. \nWe design an efficient algorithm that computes a nearly optimal non-adaptive policy for the unknown demand. \nFurthermore, we also propose a learn-then-earn policy with vanishing regret.", "candidate_comments": ["Thus motivated, we consider a model where a {\\bf pool} of unit-demand customers repeatedly interact with the seller. Each customer possesses a consistent, privately held valuation drawn independently and identically from an unknown distribution. \nWithin each time period, customers monitor the price, following a constant rate Poisson process.  Their purchasing decisions depend on the moment when they observe the price aligning with or falling below their individual valuations."]}
{"annotations": [{"id": 86958189, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86985351, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86989779, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-79afd71960e8496f.jsonl", "id_parag": "2310.19220-79afd71960e8496f_5", "final_parag": "1. {\\bf A Novel Model.} We introduce a novel {\\em pool}-based pricing model: Each customer monitors the price according to an \\indep\\ Poisson process, makes a purchase when the observed price is below the valuation, and leaves the market {\\em permanently}. \nIn contrast to the {\\em stream}-based model in most existing work, our model better encapsulates the key features of many retailing scenarios where the customers have unit demand.", "candidate_comments": ["for dynamic pricing in the context of uncertain demand. In contrast to existing literature, we consider a model where customers arrive at the start of the selling season, continuously monitor prices at a constant rate, Distinguishing our model from much of the existing literature, we assume that customers arrive at the onset of the selling season and continuously monitor the price at a constant rate. The monitoring persists until the price aligns with or falls below their individual valuation. Once a purchase decision is made, customers cease monitoring the price."]}
{"annotations": [{"id": 87254676, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87259864, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87267775, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-7a47d6b94c2d73ab.jsonl", "id_parag": "2310.19220-7a47d6b94c2d73ab_1", "final_parag": "\\begin{lemma}[Regret Decomposition] Define $\\eta_1 = |\\E[{\\rm Rev}(\\hat \\pi, \\cI')] - \\E[\\opt(\\hat \\cI)]|$ and $\\eta_2 = |\\E[\\opt(\\hat \\cI)] - \\opt(\\cI)|$. \nThen, \\[{\\rm Regret}(\\pi^{\\rm LTE}) \\leq \\eta_1 + \\eta_2.\\]", "candidate_comments": ["\\item (Optimization) $\\eta_1 := |\\E[R(\\hat s; \\cI')] - \\E[\\opt(\\hat \\cI)]|$ \\item (Interesting Name 2) $\\eta_2 := | \\E[R(\\hat s, \\hat \\cI)] - \\E[\\opt(\\hat \\cI)]|$ \\item (Estimation) $\\eta_2 := |\\E[\\opt(\\hat \\cI)] - \\opt(\\cI'')|$ \\item (Interesting Name 3) $\\eta_3 := |{\\rm OPT}({\\cal I}) -  \\E[\\opt(\\hat \\cI)]|$ \\item (Exploration) $\\eta_3 := |\\opt(\\cI'') - \\opt(\\cI)| $"]}
{"annotations": [{"id": 86983284, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986317, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86986195, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19220-7a47d6b94c2d73ab.jsonl", "id_parag": "2310.19220-7a47d6b94c2d73ab_2", "final_parag": "To this end, we first show that $\\eta_1$ will grow linearly in $s$. For this we use two observations. First, observe that our estimates are unbiased and the revenue is linear in the size of each valuation group. Second, we observe that the impact of the exploration phase (which has length $s$) is linear in $s$.", "candidate_comments": ["We first show that $\\eta_1 =0$.  This will follow because we use unbiased estimates and the error is linear in $\\hat n_i - n_i$. \nThis is essential because the error is linear in $\\hat n_i - n_i$, and we are using an unbiased estimator."]}
{"annotations": [{"id": 86980627, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86980955, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86981111, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-7a47d6b94c2d73ab.jsonl", "id_parag": "2310.19220-7a47d6b94c2d73ab_3", "final_parag": "We have $\\eta_1 = O(\\lambda n v_1 s)$. \n\\end{lemma} \\iffalse \\proof First we write $\\E[R(\\hat s; \\cI')] - \\E[\\opt(\\hat \\cI)]$ as $(\\E[R(\\hat s; \\cI')] - \\E[R(\\hat s; \\hat \\cI)]) + (\\E[R(\\hat s; \\hat \\cI)] - \\E[\\opt(\\hat \\cI)])$ and we will show that each of these terms is 0. For the first term, we condition on $\\hat s = s$ and consider the conditional expectation $\\E[R(\\hat s; \\cI') - R(\\hat s; \\hat \\cI) \\mid \\hat s = s]$. For all $s \\in [0,T]$, we have \\[ \\begin{split} R(s; \\cI') -& R(s; \\hat \\cI) = (n_1 - D - (\\hat n_1 - D))v_1 q(s) \\\\ + & (n_1 - D - (\\hat n_1 - D))v_2 (1-q(s))q(T-s) \\\\ +& (n_2 - \\hat n_2)v_2 q(T-s) \\end{split} \\] Thus the conditional expectation above is 0. Applying the law of total expectation implies the first term is 0. For the second term, this immediately follows from the definition of $\\hat s$ since it witnesses an optimal non-adaptive policy for the instance $\\hat \\cI$. \n\\endproof \\fi", "candidate_comments": ["\\proof{Proof:} We will show that $\\E[ R(\\hat s; {\\cal I}') - R(\\hat s, \\hat \\cI) \\mid \\hat s = s] =0$.  The lemma then follows from the law of total expectation.  The only way the instances $\\hat \\cI$ and $\\cI'$ differ is in the number of type 1 and type 2 customers.  Using this we observe that for any fixed $s \\in [0,T]$, we have \\[ \\begin{split}     R(s; {\\cal I}') -  R(s, \\hat \\cI)  = &  (n_1 - D - (\\hat n_1 - D))v_1 q(s) \\\\       & + (n_1 - D - (\\hat n_1 - D))v_2 (1-q(s))q(T-s) + (n_2 - \\hat n_2)v_2 q(T-s) \\\\       = & (n_1 - \\hat n_1)(v_1q(s) + v_2(1-q(s))q(T-s)) + (n_2-\\hat n_2)v_2q(T-s) \\end{split} \\] Therefore taking the expectation conditional on $\\hat s = s$, we have \\[\\E[ R(\\hat s; {\\cal I}') - R(\\hat s, \\hat \\cI) \\mid \\hat s = s] = 0\\] since $\\hat n_1$ and $\\hat n_2$ are unbiased estimates of $n_1$ and $n_2$. \n% \\hfill\\Halmos  \\endproof"]}
{"annotations": [{"id": 86962225, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973010, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86973414, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-7a47d6b94c2d73ab.jsonl", "id_parag": "2310.19220-7a47d6b94c2d73ab_4", "final_parag": "The following is a standard concentration inequality for binomial random variables, see, e.g., Theorem 2.26 in~\\cite{VershyninHDP}. \nWe will use this to show that the estimates $\\hat n_i$ we compute in the learning phase are likely to be close to the true values $n_i$.", "candidate_comments": ["Under this condition we can show that the optimal revenue on $\\hat \\cI$ is close to the optimal revenue on $\\cI$.   First we state the lemma and then develop some tools to prove it."]}
{"annotations": [{"id": 86722084, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86726167, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86727317, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-7cf68fd41234e038.jsonl", "id_parag": "2310.19220-7cf68fd41234e038_1", "final_parag": "\\subsection{Upper Bounding on the Optimal Revenue} An alternative idea is to find a closed-form formula for the denominator for any given $(n_i)$'s, and reduce the bilevel problem to a single-level problem.", "candidate_comments": ["TL: Cutting next sentence since we redefined OPT. \nIn fact, it is quite challenging to even derive a closed-form expression for the revenue of an optimal  {\\em non-\\adap} policy, given the intricate nature of  the revenue \\func\\ in \\ref{prop:exp_rev}."]}
{"annotations": [{"id": 86738675, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86902107, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86903718, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19220-7cf68fd41234e038.jsonl", "id_parag": "2310.19220-7cf68fd41234e038_2", "final_parag": "To see this, note that if a customer has valuation $v$, then the maximum expected revenue from this customer is at most $v(1-e^{-\\lambda})$, which is attained by the policy that always selects price $v$. The expression ${\\rm UB}(\\mathcal{I})$ is simply the sum of this upper bound over all customers.", "candidate_comments": ["the seller offers them a price exactly as their valuation $v_i$. \nNote that $v_i (1-e^{-\\lambda T})$ is the largest revenue the seller can extract from customers with valuation $v_i$, which is the product of the highest price a customer with valuation $v_i$ will accept and the monitoring probability in time $T$. \nNo other price policy can achieve higher revenue. \nTherefore, ${\\rm UB}(\\mathcal{I})$ is an upper bound for any (possibly adaptive) policy."]}
{"annotations": [{"id": 86985974, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86987017, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86989993, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-7cf68fd41234e038.jsonl", "id_parag": "2310.19220-7cf68fd41234e038_3", "final_parag": "On the other hand, it should be noted that ${\\rm UB}(\\mathcal{I})$ can be much greater than ${\\rm OPT}(\\mathcal{I})$. In fact, the ${\\rm UB}(\\mathcal{I})$ is attained by a {\\em personalized} policy, i.e., prices for different customers may differ, whereas ${\\rm OPT}(\\mathcal{I})$ is defined over {\\em non-personalized} policies.", "candidate_comments": ["the seller offers them a price exactly as their valuation $v_i$. \nNote that $v_i (1-e^{-\\lambda T})$ is the largest revenue the seller can extract from customers with valuation $v_i$, which is the product of the highest price a customer with valuation $v_i$ will accept and the monitoring probability in time $T$. \nNo other price policy can achieve higher revenue. \nTherefore, ${\\rm UB}(\\mathcal{I})$ is an upper bound for any (possibly adaptive) policy."]}
{"annotations": [{"id": 86725887, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86728260, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86729872, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-7cf68fd41234e038.jsonl", "id_parag": "2310.19220-7cf68fd41234e038_4", "final_parag": "\\subsection{Linearization} With this upper bound, we next focus on the bilevel optimization problem where ${\\rm OPT}(\\mathcal{I})$ is replaced with ${\\rm UB}(\\mathcal{I})$. Explicitly, we consider The above bilevel problem is still not readily solvable since ${\\rm Rev}(\\pi, \\mathcal{I})$ and ${\\rm UB}(\\mathcal{I})$ are both {\\em non-linear} \\func s. \nTo circumvent this, we consider a {\\em linear surrogate} for each of them, motivated by Taylor's expansion.", "candidate_comments": ["We refer to ${\\rm Rev}'(\\pi, \\mathcal{I})$ and $UB'(\\mathcal{I})$ as the surrogate ${\\rm Rev}(\\pi, \\mathcal{I})$ and $UB(\\mathcal{I})$."]}
{"annotations": [{"id": 86862167, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86901313, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86901455, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-7cf68fd41234e038.jsonl", "id_parag": "2310.19220-7cf68fd41234e038_5", "final_parag": "\\bdefn[Linear Surrogate] For any \\ins\\ $\\mathcal{I} =(\\lambda, \\{n_i\\}_{i\\in [k]}, \\{v_i\\}_{i\\in [k]})$ and non-\\adap\\ policy $\\pi = (t_i)$, we define the linear surrogate of ${\\rm UB(\\mathcal{I})}$ and ${\\rm Rev(\\pi, \\mathcal{I})}$ as \\edefn", "candidate_comments": ["We refer to ${\\rm Rev}'(\\pi, \\mathcal{I})$ and $UB'(\\mathcal{I})$ as the surrogate ${\\rm Rev}(\\pi, \\mathcal{I})$ and $UB(\\mathcal{I})$."]}
{"annotations": [{"id": 86981085, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86986029, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86986182, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-7cf68fd41234e038.jsonl", "id_parag": "2310.19220-7cf68fd41234e038_6", "final_parag": "For any instance $\\mathcal{I} = (\\lambda,\\{n_i\\}_{i=1}^k,\\{v_i\\}_{i=1}^k)$ and non-\\adap\\ policy $\\pi$, we have \\[\\frac{{\\rm Rev}(\\pi, \\mathcal{I})}{{\\rm UB}(\\mathcal{I})} \\ge \\frac{{\\rm Rev'}(\\pi, \\mathcal{I})}{{\\rm UB'}(\\mathcal{I})}.\\] \\end{lemma}", "candidate_comments": ["We refer to ${\\rm Rev}'(\\pi, \\mathcal{I})$ and $UB'(\\mathcal{I})$ as the surrogate ${\\rm Rev}(\\pi, \\mathcal{I})$ and $UB(\\mathcal{I})$."]}
{"annotations": [{"id": 86964386, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86972598, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86973455, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-7cf68fd41234e038.jsonl", "id_parag": "2310.19220-7cf68fd41234e038_7", "final_parag": "$(1-e^{-x})/x \\ge (1-e^{-y})/y$, which rearranges to \\[\\frac{1 - e^{-x}}{1 - e^{-y}} \\ge \\frac xy.\\]", "candidate_comments": ["\\[\\frac{1 - e^{-x}}{x} \\ge \\frac{1 - e^{-y}}{y},\\] i.e., \\[\\frac{1 - e^{-x}}{1 - e^{-y}} \\ge \\frac xy.\\]"]}
{"annotations": [{"id": 86725340, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86731606, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86730334, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-7cf68fd41234e038.jsonl", "id_parag": "2310.19220-7cf68fd41234e038_8", "final_parag": "\\begin{theorem}[Competitive Ratio Lower Bound] For any instance $\\mathcal{I} = (\\lambda, \\{n_i\\}_{i=1}^k,\\{v_i\\}_{i=1}^k)$ where $\\{n_i\\}_{i=1}^k$ is unknown to the seller, we can compute in polynomial time a non\\adap\\ policy ${\\bf \\pi } = (t_1, \\cdots, t_{k+1})$ such that \\[\\frac{{\\rm Rev}(\\pi, \\mathcal{I})}{{\\rm OPT}(\\mathcal{I})} \\ge \\frac{1}{1 + \\log(v_1/v_k)}.\\] \\end{theorem} Since the optimal revenue on the whole valuation set, ${\\rm OPT}(\\mathcal{I})$, is at most $(1 + \\epsilon)$ of the optimal revenue on the core set, the competitive ratio we derive on the core set is at least, \\[ \\frac{1}{(1 + \\epsilon)(k - \\sum_{i=1}^{k-1} v_{i+1}/v_{i})}, \\] where $k = \\log(v_1/v_k)/\\log(1+\\eps)$ for the core set, and $v_{i+1}/v_{i} = 1/(1 + \\epsilon)$ for $i \\in [k-1]$. Plugging the expression of $k$ and $v_{i+1}/v_{i} $, then the competitive ratio on the core is \\[ \\frac{1}{1 + \\epsilon \\log(v_1/v_k)/\\log(1+\\eps)} . \n\\] Note $\\epsilon/\\log(1+\\eps)$ is increasing in $\\epsilon$ and $\\epsilon/\\log(1+\\eps) = 1$ when $\\epsilon$ goes to 0, therefore, the competitive ratio is at least, \\[ \\frac{1}{1 + \\epsilon \\log(v_1/v_k)/\\log(1+\\eps)} \\ge \\frac{1}{1 + \\log(v_1/v_k)} . \n\\]", "candidate_comments": ["Using the previous $1/k$ result, the competitive ratio will be at least $1/((1 + \\epsilon)\\log(b/a)/\\log(1+\\epsilon))$."]}
{"annotations": [{"id": 86952223, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86959034, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86959513, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19220-7cf68fd41234e038.jsonl", "id_parag": "2310.19220-7cf68fd41234e038_9", "final_parag": "We also show that the above lower bound of $1/k$ is the best possible. \nNo algorithm can achieve a fraction larger than $1/k$ of the optimal revenue. \nIn \\cref{thm:non_adap_upper}, we demonstrate that for any non-\\adap\\ policy $\\pi$, there exists an instance such that the policy can achieve at most $\\frac{1}{k} + \\epsilon$ fraction of the optimal revenue.", "candidate_comments": ["on which the output of this policy earns at most $\\frac{1}{k} + \\eps$ fraction of the optimal revenue."]}
{"annotations": [{"id": 86982701, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86984470, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86984684, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19220-9faee767d205059f.jsonl", "id_parag": "2310.19220-9faee767d205059f_1", "final_parag": "A compelling class of policies is the class of non-\\adap\\ policies, where the prices are determined upfront, regardless of the purchase events. \nThese policies are widely applied in practice due to their simplicity and effectiveness; see, e.g., \\cite{ma2021dynamic}.", "candidate_comments": ["In this section, we consider {\\it non-\\adap\\ policies}, i.e., policies where the price sequence and time at each price are determined in advance, regardless of the realized demands. We formally define it as follows."]}
{"annotations": [{"id": 86961951, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973104, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86973907, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19220-9faee767d205059f.jsonl", "id_parag": "2310.19220-9faee767d205059f_2", "final_parag": "We will therefore restrict our attention to non-\\adap\\ markdown policies subsequently. Each policy in this class can be specified by a sequence $(t_1,\\cdots, t_k)$ where the policy selects the price $v_i$ from time $t_i$ to $t_{i+1}$.", "candidate_comments": ["We establish the expected revenue expression for any non-adaptive markdown policy.  Recall that for each $i \\in [k]$, we denote $R_i$ as the (random) revenue generated from a customer with valuation $v_i$.\\su{OK}"]}
{"annotations": [{"id": 86983405, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86985520, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86985607, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19232-65bdbd0d341cfb6c.jsonl", "id_parag": "2310.19232-65bdbd0d341cfb6c_1", "final_parag": "\\paragraph{Notations used:} Henceforth, we denote $\\bm{W}_d,\\; \\bm{W}_u,\\;\\bm{h}$ by $\\mathbf{A},\\;\\mathbf{B},\\;\\text{and}\\;\\mathbf{x}$, respectively; $\\mathbf{B}^+ {\\coloneqq} \\max \\{\\mathbf{B}, \\mathbf{0}\\}$; $\\mathbf{B}^- {\\coloneqq} \\max \\{-\\mathbf{B}, \\mathbf{0}\\}$; $b_i$ denotes $i_{th}$ row of $\\mathbf{B}$; $\\bm{b}^{i+}{\\coloneqq}\\max \\{\\bm{b}^i,\\bm{0}\\},$ $\\bm{b}^{i-}{\\coloneqq}\\max \\{-\\bm{b^i},\\bm{0}\\}$; $\\diag[\\bm{u}]$ arranges $\\bm{u}$ in a diagonal matrix; $||\\mathbf{G}||_{1,1} {\\coloneqq} \\Sigma_{k=1}^d ||\\mathbf{G}(i,:)||_1$; $||\\cdot||_F$ denotes Frobenius Norm.", "candidate_comments": ["\\paragraph{Notations used:} We denote: \\begin{itemize}     \\item $\\bm{W}_d$, $\\bm{W}_u$, $\\bm{h}$ by $\\mathbf{A}$, respectively,     \\item $\\mathbf{B}^+  \\coloneqq \\max \\{\\mathbf{B}, \\mathbf{0}\\}$,     \\item $\\bm{b}^{i+}{\\coloneqq}\\max \\{\\bm{b}^i,\\bm{0}\\},$ $\\bm{b}^{i-}{\\coloneqq}\\max \\{-\\bm{b^i},\\bm{0}\\}$,     \\item $\\diag[\\bm{u}]$ arranges $\\bm{u}$ in a diagonal matrix,     \\item $||\\mathbf{G}||_{1,1} \\coloneqq \\Sigma_{k=1}^d ||\\mathbf{G}(i,:)||_1$     \\item $||\\cdot||_F$ denotes Frobenius Norm \\end{itemize}"]}
{"annotations": [{"id": 86904639, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86981071, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86981237, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19232-65bdbd0d341cfb6c.jsonl", "id_parag": "2310.19232-65bdbd0d341cfb6c_2", "final_parag": "Thus, $f(\\mathbf{x})=H(\\mathbf{x})-Q(\\mathbf{x})$. Let $f^i$ denote the first output from adapter block, $b^i=\\mathbf{B}[i,:]$ (i.e. $i^{\\text{th}}$ row of $\\mathbf{B}$). We use $\\mathcal{H}$ and $\\mathcal{Q}$ to denote tropical hypersurfaces of $H$ and $Q$ at node $i$. \nComputing dual subdivision", "candidate_comments": ["Let $f^i$ denote the first output from adapter block, following this, we define $f^i = h^i-q^i$, $b^i=\\mathbf{B}[i,:]$ (i.e. $i^{\\text{th}}$ row of $\\mathbf{B}$). Tropically we can denote $h^i$ as"]}
{"annotations": [{"id": 87254824, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87258576, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87268100, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19233-65bdbd0d341cfb6c.jsonl", "id_parag": "2310.19233-65bdbd0d341cfb6c_1", "final_parag": "This paper studies how to effectively build meeting summarization systems for real-world usage using large language models (LLMs). For this purpose, we conduct an extensive evaluation and comparison of various closed-source and open-source LLMs, namely, GPT-4, GPT-3.5, PaLM-2, and LLaMA-2. Our findings reveal that most closed-source LLMs are generally better in terms of performance. However, much smaller open-source models like LLaMA-2 (7B and 13B) could still achieve performance comparable to the large closed-source models even in zero-shot scenarios. Considering the privacy concerns of closed-source models for only being accessible via API, alongside the high cost associated with using fine-tuned versions of the closed-source models, the open-source models that can achieve competitive performance are more advantageous for industrial use. Balancing performance with associated costs and privacy concerns, the LLaMA-2-7B model looks more promising for industrial usage. \nIn sum, this paper offers practical insights on using LLMs for real-world business meeting summarization, shedding light on the trade-offs between performance and cost.", "candidate_comments": ["In this paper, we present an evaluation of commercial and open-source Large Language Models (LLMs) for automated meeting summarization. We leverage the MeetingBank dataset to benchmark two commercial LLMs - OpenAI's GPT-3.5 and Google's PaLM-2 - against three open-source LLMs - RedPajama, MosaicML, and Falcon. Our findings reveal superior performance by commercial LLMs in producing accurate and coherent meeting summaries, although the open-source alternatives also display potential. Notably, the usage of open-source LLMs incurs a higher computational GPU cost. These results contribute to our understanding of LLM efficacy in extended meeting summarization, underscoring the cost-effectiveness of commercial options versus the computational expense of open-source models. The insights provided here are poised to aid industry and academia in creating efficient and economical meeting summarization systems. The paper hence offers guidance in the selection and application of LLMs for real-world use-cases. %or not based on the quality assurance (QA) queries/questions asked by the contact center managers and"]}
{"annotations": [{"id": 86868515, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86904272, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86909940, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19233-65bdbd0d341cfb6c.jsonl", "id_parag": "2310.19233-65bdbd0d341cfb6c_2", "final_parag": "(iii) Finally, we demonstrate a practical perspective on the trade-offs that come with selecting a model for real-world usage based on its performance, cost, and computational requirements.", "candidate_comments": ["his paper thus presents a detailed examination of the use of LLMs for the specific task of meeting summarization, offering a practical perspective on the trade-offs that come with selecting a model based on its performance, cost, and computational requirements. The results and insights we share in this paper serve to inform and guide researchers, practitioners, and industry professionals alike in their quest to harness the power of LLMs for efficient meeting summarization."]}
{"annotations": [{"id": 86984852, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986157, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86986208, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19233-65bdbd0d341cfb6c.jsonl", "id_parag": "2310.19233-65bdbd0d341cfb6c_3", "final_parag": "\\textbf{PaLM-2:} PaLM-2 is an LLM \\cite{palm2} developed by Google. It leverages the mixture of objectives technique \\cite{palm2} and significantly outperforms the original PaLM \\cite{chowdhery2022palm} model. We use the \\textit{text-bison@001 model in \\textit{Google's VertexAI\\footnote{\\url{https://cloud.google.com/vertex-ai/docs/generative-ai/model-reference/text}}}} for PaLM-2 that has an input context window of 8K tokens.", "candidate_comments": ["It is a Transformer decoder-only LLM developed by Google that leverages the mixture of objectives technique \\cite{palm2}. It achieved impressive performance on several downstream tasks while being more efficient in terms of the inference speed compared to the original PaLM \\cite{chowdhery2022palm} model. The PaLM-2 model has an input context window of 8K tokens. In our experiments, we use the \\textit{text-bison@001 model in \\textit{Google's VertexAI\\footnote{\\url{https://cloud.google.com/vertex-ai/docs/generative-ai/model-reference/text}}}} for PaLM-2.  %Thus, to ensure a fair comparison with GPT-3.5 and GPT-4, we experiment with the token size of ($n=2500$) and ($n=5000$) for truncation and chapterization. "]}
{"annotations": [{"id": 86963130, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86976099, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86976575, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19233-65bdbd0d341cfb6c.jsonl", "id_parag": "2310.19233-65bdbd0d341cfb6c_4", "final_parag": "(vi) In the QMSUM dataset, the \\textit{truncation} approach is found to be the best-performing approach, as GPT-4 using this technique achieves the best result in terms of ROUGE-1\\&2, while GPT-3.5 achieves the best based on ROUGE-L. However, in terms of BERTScore, we find that GPT-4 based on \\textit{chapterization via resummarization} approach to be the best. These findings further establish our intuition that in datasets where the average gold reference summary length is smaller (in QMSUM, the average summary length is just 108 words), \\textit{chapterization via resummarization} or \\textit{summarization via truncation} approaches are more useful.", "candidate_comments": ["Meanwhile, the \\textit{summarization via truncation} approach proves to be cost-efficient.   (iv) We observe a significant drop in performance for PaLM-2 using \\textit{summarization via truncation} and \\textit{chapterization via resummarization} approaches in the QMSUM dataset. However, the performance was better in GPT-3.5 and GPT-4 using these approaches. This may indicate that summarizing based on less contextual information yields poorer results for PaLM-2. "]}
{"annotations": [{"id": 86971381, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86973212, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86974634, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19233-65bdbd0d341cfb6c.jsonl", "id_parag": "2310.19233-65bdbd0d341cfb6c_5", "final_parag": "(viii) Based on the average across all datasets in terms of different summarization types (see Figure \\ref{fig:average_results_types}), the \\textit{chapterization via rewriting} approach is found to be the best. We also do not find any significant difference in performance based on various summarization types. Since the truncation-based approach achieves performance comparable to various chapterization-based approaches even without considering the whole context, further investigation of the gold summaries in these datasets is needed.", "candidate_comments": [" (iv) We observe a significant drop in performance for PaLM-2 using \\textit{summarization via truncation} and \\textit{chapterization via resummarization} approaches in the QMSUM dataset. However, the performance was better in GPT-3.5 and GPT-4 using these approaches. This may indicate that summarizing based on less contextual information yields poorer results for PaLM-2. "]}
{"annotations": [{"id": 86981700, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86984200, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86984434, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19233-65bdbd0d341cfb6c.jsonl", "id_parag": "2310.19233-65bdbd0d341cfb6c_6", "final_parag": "To further investigate the performance of LLMs, we conduct some case studies using the \\textit{Summarization via Truncation} approach, as demonstrated below.", "candidate_comments": ["However, we do not use it since OpenAI announces that \\textit{text-da-vinci-003} will be deprecated soon."]}
{"annotations": [{"id": 87254926, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87258922, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87268050, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19233-65bdbd0d341cfb6c.jsonl", "id_parag": "2310.19233-65bdbd0d341cfb6c_7", "final_parag": "\\textbf{Inference Speed:} We also measure the inference speed of different LLMs. For this purpose, we collected 100 meeting transcripts from Dialpad\\footnote{\\url{https://www.dialpad.com/ca/}} that consist of real-world conversations. All the collected transcripts have at least 2500 words. Based on our analysis using the \\textit{Summarization via Truncation} approach, we find that GPT-3.5 is the fastest, as it takes 2.5 seconds on average per transcript for inference, followed by PaLM-2 which takes about 3.2 seconds on average. While GPT-4 achieves the best performance in terms of ROUGE and BERTScore metrics, it is also the slowest among the commercial closed-source LLMs since it takes about 11 seconds on average per transcript. We also benchmark the inference speed of LLaMA-2-7B in GCP on a machine having 1 NVIDIA L4 GPU and find that it takes 15 seconds on average.", "candidate_comments": ["To measure the inference speed of different closed-source LLMs, we use  (only Test) version of the QMSUM\\textsubscript{Filtered} dataset. It contains 37 samples where the average length of the transcript is about 10546 words. Based on our analysis, we find that GPT-3.5 is the fastest, as it takes 3 seconds per transcript for inference, whereas PaLM-2 takes 8 seconds per transcript and GPT-4 takes 40 seconds per transcript.  This is a section in the appendix."]}
{"annotations": [{"id": 86904644, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973266, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86974783, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19250-79b6f17d20278492.jsonl", "id_parag": "2310.19250-79b6f17d20278492_1", "final_parag": "\\item There is no published work that systematically studies the utility and fairness of machine learning models trained on several GAN based and marginal-based synthetic tabular data set generation algorithms.", "candidate_comments": ["About Fairness In domains such as health care, education and consumer finance, the usage of machine learning models trained on sensitive data is ubiquitous \\cite{chen2017disease,ghoddusi2019machine}. "]}
{"annotations": [{"id": 86962805, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86971871, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86972122, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19250-79b6f17d20278492.jsonl", "id_parag": "2310.19250-79b6f17d20278492_2", "final_parag": "\\item Previous studies have not evaluated machine learning models trained on synthetic data set generation algorithms for multiple definitions of fairness.", "candidate_comments": ["About Fairness In domains such as health care, education and consumer finance, the usage of machine learning models trained on sensitive data is ubiquitous \\cite{chen2017disease,ghoddusi2019machine}. "]}
{"annotations": [{"id": 86961627, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86963066, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86965668, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19250-79b6f17d20278492.jsonl", "id_parag": "2310.19250-79b6f17d20278492_3", "final_parag": "The promises synthetic data brings generated an interest in understanding impacts of utilizing synthetic in data analysis and machine learning. Some of these works include analysing the utility of differentially private synthetic data in different tasks \\cite{tao2021benchmarking}, investigating if training models with differentially private synthetic images can increase subgroup disparities \\cite{cheng2021can}, the impacts different types of synthetic data can have in model fairness \\cite{ganev2022robin, bullwinkel2022evaluating}, utility of synthetic data in downstream health care classification systems \\cite{movahedi2023evaluating}, and whether feature importance can be accurately analyzed using differentially private synthetic data \\cite{giles2022faking}. All these works are ultimately trying to answer a same question: to which extent can we substitute real data with synthetic data, and which are the best synthetic data generation techniques for model training?", "candidate_comments": ["Despite For instance, it is not known how different synthetic data generation methods affect fairness. Also, in the literature, it is usually assumed that real data is available for testing models trained on synthetic data prior to deployment. This is not a realistic assumption if the goal is to democratize data and make more data publicly available. In such case, only synthetic data is available during training and testing. So, it is important to study the performance of a model trained and tested with synthetic data vs its performance when tested against real data. In this paper we address these questions.  "]}
{"annotations": [{"id": 86961098, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86965093, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86965254, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19250-79b6f17d20278492.jsonl", "id_parag": "2310.19250-79b6f17d20278492_4", "final_parag": "These findings highlight synthetic data's potential reliability and viability as a substitute for real data sets in end-to-end machine learning pipelines for tabular data. Furthermore, our research sheds light on the implications of model fairness when utilizing differentially private synthetic data for model training.", "candidate_comments": ["The real data sets in our experiments are not large in size ($\\approx$30k samples Adult; $\\approx$8k samples COMPAS) for big data standards and have low dimensionality.  We leave as future work a similar investigation with other synthesizer techniques and data sets with larger sample sizes and high dimensionality. \nAdditionally, our experiments reveal the necessity of model testing protocols in scenarios where differentiallt private synthetic data sets are utilized for model training and evaluation."]}
{"annotations": [{"id": 86737789, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86961185, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86959632, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19272-f91e66020a25c41e.jsonl", "id_parag": "2310.19272-f91e66020a25c41e_1", "final_parag": "\\paragraph{Incompatibility with logits-based replay.} The NPCL is incompatible with logits-based replay because of the stochasticity in the posterior induced by Monte Carlo sampling. Overcoming this could help boost the performance of the NPCL further over SOTA like DER \\cite{NEURIPS2020_b704ea2c} and DER++ \\cite{boschini2022class}.", "candidate_comments": ["Future works will address the following two limitations. First, the NPCL is incompatible with logits-based replay because of the stochasticity in the posterior induced by  Monte Carlo sampling. Second, the posterior approximation gap \\cite{anonymous2023bridge} and the attention mechanism (App. \\ref{subsec:ca_analysis}) limits the sensitivity of the NPCL to increased context samples. "]}
{"annotations": [{"id": 87254938, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}, {"id": 87259095, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}, {"id": 87267663, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19273-eb1b6e6dd2633b29.jsonl", "id_parag": "2310.19273-eb1b6e6dd2633b29_1", "final_parag": "\\setlength{\\parsep}{\\parskip} \\setlength{\\itemsep}{0cm} \\setlength{\\topsep}{\\parskip} \\setlength{\\parskip}{0cm} \\setlength{\\partopsep}{0cm} \\setlength{\\listparindent}{\\parindent} \\setlength{\\labelwidth}{10pt} \\setlength{\\labelsep}{0pt} \\setlength{\\leftskip}{0pt} \\setlength{\\leftmargin}{0pt} }", "candidate_comments": ["change margins \\setlength{\\textwidth}{7in} \\setlength{\\textheight}{8.75in} \\setlength{\\oddsidemargin}{-0.25in} \\setlength{\\evensidemargin}{-0.25in} \\setlength{\\headsep}{10pt}", "Latex \\documentstyle[fleqn,psfig,epsfig]{article} \\documentstyle[psfig]{article} \\setlength{\\textwidth}{6.5in} \\setlength{\\oddsidemargin}{0in} \\setlength{\\textheight}{8.5in} \\setlength{\\headheight}{0in} \\setlength{\\headsep}{0in} \\setlength{\\parindent}{0in} % block style \\setlength{\\parskip}{0.3cm}"]}
{"annotations": [{"id": 86734354, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86959160, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86959533, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19273-eb1b6e6dd2633b29.jsonl", "id_parag": "2310.19273-eb1b6e6dd2633b29_2", "final_parag": "\\newtheorem{claim}{Claim}{} \\newtheorem{thm}{Theorem}{} \\newtheorem{prop}{Proposition}{} \\newtheorem{corr}{Corollary}{}", "candidate_comments": ["\\newtheorem{remark}{Remark} \\newtheorem{lemma}{Lemma}{} \\newtheorem{assump}{Assumption} \\newtheorem{defn}{Definition}{}"]}
{"annotations": [{"id": 86984395, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86985917, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86985891, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19273-eb1b6e6dd2633b29.jsonl", "id_parag": "2310.19273-eb1b6e6dd2633b29_3", "final_parag": "\\newenvironment{mythm}{{\\bf Theorem}}{} \\newenvironment{myproof}{{\\bf Proof:}}{}", "candidate_comments": ["We want to avoid name clashes with the exercise.sty package \\newenvironment{myexercises}{\\section{Exercises}}{}"]}
{"annotations": [{"id": 86981543, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86983047, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86983948, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19273-eb1b6e6dd2633b29.jsonl", "id_parag": "2310.19273-eb1b6e6dd2633b29_4", "final_parag": "\\newenvironment{myexercises}{\\section{Exercises}}{}", "candidate_comments": ["We want to avoid name clashes with the exercise.sty package \\newenvironment{myexercises}{\\section{Exercises}}{}"]}
{"annotations": [{"id": 86952259, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86958390, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86960566, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19273-eb1b6e6dd2633b29.jsonl", "id_parag": "2310.19273-eb1b6e6dd2633b29_5", "final_parag": "\\newenvironment{mysolutions}{\\section{Solutions}}{}", "candidate_comments": ["We want to avoid name clashes with the exercise.sty package \\newenvironment{myexercises}{\\section{Exercises}}{}"]}
{"annotations": [{"id": 86972416, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86982683, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86982576, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19274-673bce1930413ea6.jsonl", "id_parag": "2310.19274-673bce1930413ea6_1", "final_parag": "This is accomplished by partitioning the lower-dimensional space $\\bm{\\chi}$ into overlapping subsets, ${u_i}$, $i=1, \\cdots, N$, and $N$ is the total number of subsets.", "candidate_comments": ["Next, we perform a covering stage, which partitions the lower-dimensional space $\\bm{\\chi}$ into overlapping subsets. This can be mathematically represented as ${U_i}_{i=1}^{N}$, where each $U_i$ is a subset of $\\bm{\\chi}$ and $N$ is the total number of subsets. "]}
{"annotations": [{"id": 86982711, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86983500, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86983918, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19274-ffcf0f83f3949d72.jsonl", "id_parag": "2310.19274-ffcf0f83f3949d72_1", "final_parag": "\\subsection{Closeness Centrality} Closeness Centrality of a vertex \\(v\\) measures how close the vertex is to all other vertices in the graph, defined as the inverse of the sum of the shortest path distances from \\(v\\) to all other vertices. Given a graph \\(G(V, E)\\), the closeness centrality \\(C_C(v)\\) of a vertex \\(v\\) can be computed as in equation \\ref{eqn:closeness}.", "candidate_comments": ["\\subsection{Betweenness Centrality} Betweenness Centrality of a vertex $v$ is a measure of the influence of a node in the network, defined as the fraction of the shortest paths that pass through $v$. More formally, given a graph $G(V, E)$, betweenness centrality $C_B(v)$ for a vertex $v$ can be computed as shown in equation \\ref{eqn:betweenness}."]}
{"annotations": [{"id": 86738021, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86765551, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86827496, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19274-ffcf0f83f3949d72.jsonl", "id_parag": "2310.19274-ffcf0f83f3949d72_2", "final_parag": "where \\(d(u, v)\\) is the shortest path distance between vertices \\(u\\) and \\(v\\).", "candidate_comments": ["where $\\sigma_{st}$ is the total number of shortest paths from node $s$ to node $t$ and $\\sigma_{st}(v)$ is the number of those paths that pass through $v$."]}
{"annotations": [{"id": 86899058, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86899809, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86908450, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19275-6daba0ea296e90c3.jsonl", "id_parag": "2310.19275-6daba0ea296e90c3_1", "final_parag": "We then tested three different types of LLM prompting techniques to see which would best recreate a Wikipedia topic hierarchy for subtopics of Computer Science\\footnote{https://en.wikipedia.org/wiki/Category:Computer\\_science}''. To evaluate the LLM outputs, we developed evaluation criteria to assess the accuracy of the prompt outputs based on specificity (is the output too broad or too specific?), lack of repetitiveness (does the outputs regurgitate the topic that was inputted?), and relatedness (is the output unrelated or tangentially related?).", "candidate_comments": ["We used the topics in Wikipedia's ``Category of Computer Science\\footnote{https://en.wikipedia.org/wiki/Category:Computer\\_science}'' to extract topic information and create a hierarchical topic structure with five subtopic levels from most general to most specific. "]}
{"annotations": [{"id": 86862899, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86901808, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86903696, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19285-66c30bd111f2a070.jsonl", "id_parag": "2310.19285-66c30bd111f2a070_1", "final_parag": "Node-level random walk has been widely used to improve Graph Neural Networks. However, there is limited attention to random walk on edge and, more generally, on $k$-simplices. This paper systematically analyzes how random walk on different orders of simplicial complexes (SC) facilitates GNNs in their theoretical expressivity. First, on $0$-simplices or node level, we establish a connection between existing positional encoding (PE) and structure encoding (SE) methods through the bridge of random walk. Second, on $1$-simplices or edge level, we bridge edge-level random walk and Hodge $1$-Laplacians and design corresponding edge PE respectively. In the spatial domain, we directly make use of edge level random walk to construct EdgeRWSE. Based on the spectral analysis of Hodge $1$-Laplcians, we propose Hodge1Lap, a permutation equivariant and expressive edge-level positional encoding. Third, we generalize our theory to random walk on higher-order simplices and propose the general principle to design PE on simplices based on random walk and Hodge Laplacians. Inter-level random walk is also introduced to unify a wide range of simplicial networks. Extensive experiments verify the effectiveness of our random walk-based methods. \n\\end{abstract}", "candidate_comments": ["Node-level random walk on graphs has been widely studied to improve the expressivity or representation learning of Graph Neural Networks. However, there are limited attention to random walk on edge-level, and more generally, on $k$-simplicials. In this paper, we systematically analyze how random walk on different orders of simplicials facilitates GNNs in their theoretical expressivity and representation power. First, on $0$-simplicials or node level, we establish a connection between subgraph sampling and edge dropping methods through the bridge of random walk. We further introduce distance metrics to node-level random walk and propose a novel probabilistic message passing mechanism, which can jointly learn features and structures of nodes. Positional encoding methods based on node-level random walk are also theoretically studied and compared. Second, on $1$-simplicials or edge level, discrete Hodge-deRham theory is applied to bridge edge-level random walk and Hodge 1-Laplacians. Based on spectral analysis of Hodge 1-Laplcians, we propose Hodge1Lap, the first equivariant edge-level positional encoding that contains rich structure information. Third, we use random walk on higher-order or inter-order simplicials to unify a wide range of simplicial networks, and generalize the techniques to learning on graphs other than on simplicials. Extensive experiments verify the effectiveness of our random walk based methods, including probabilistic message passing and Hodge1Lap positional encoding."]}
{"annotations": [{"id": 86961171, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86972405, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86972376, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19285-66c30bd111f2a070.jsonl", "id_parag": "2310.19285-66c30bd111f2a070_2", "final_parag": "\\begin{theorem} RWSE is strictly less powerful than $2$-FWL, i.e. RWSE $\\prec$ $2$-FWL. \n\\end{theorem}", "candidate_comments": ["\\begin{theorem}     PE and SE as injective function of $\\mathbf P$ and its powers are not more powerful than $2$-FWL. Specifically, RWSE is strictly less powerful than $2$-FWL. \n\\end{theorem}"]}
{"annotations": [{"id": 86736877, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86902263, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86903562, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19285-66c30bd111f2a070.jsonl", "id_parag": "2310.19285-66c30bd111f2a070_3", "final_parag": "Graph Laplacian, or Hodge $0$-Laplacian as we refer to later, is closely connected with random walk on graph. The definition of graph Laplacian is $\\mathbf L_0=\\mathbf D-\\mathbf A=\\delta_0^*\\delta_0=\\Delta_0$. Through the spectrum of $\\mathbf L_0$, we are able to define a family of positive definite kernels on graphs \\citep{KernelsAR} by applying a regularization function $r$ to the spectrum of $\\mathcal L_0$: $K_r=\\sum_{i=1}^m r(\\lambda_i) \\mathbf u_i \\mathbf u_i^T$, where $\\mathbf L_0=\\sum_i\\lambda_i \\mathbf u_i\\mathbf u_i^T$ is the eigenvalue decomposition. For example, the heat kernel or the diffusion kernel \\citep{WLInfiniteheatkernel} can be incorporated if $r(\\lambda_i)=e^{-\\beta \\lambda_i}$. Other methods directly use eigenvectors as PE \\citep{RethinkingGTLap}. These results imply that spectral analysis of graph Laplacians can also inspire more powerful PE and SE, and we will generalize graph Laplacian $\\mathbf L_0$ to arbitrary order of Hodge $k$ Laplacians in the following section to facilitate graph learning.", "candidate_comments": ["Graph Laplacian, or Hodge $0$-Laplacian as we refer to later, is closely connected with random walk on graph. The definition of graph Laplacian is $\\mathbf L_0=\\mathbf D-\\mathbf A=\\delta_0^*\\delta_0=\\Delta_0$. A well known random walk normalization form of $\\mathbf L_0$ is $\\mathbf L_0^{RW}=\\mathbf D^{-1}\\mathbf L_0=\\mathbf I-\\mathbf D^{-1}A$. \\cai{further discussion on random walk and stability of eigen-decomposition}"]}
{"annotations": [{"id": 86952647, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86961072, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86959662, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19285-66c30bd111f2a070.jsonl", "id_parag": "2310.19285-66c30bd111f2a070_4", "final_parag": "The two variations of edge-level random walk via down-neighbors differ in whether two lower adjacent nodes of the edge have the same status. Concretely, the first type of edge-level random walk based on $\\mathbf P_{down}$, which we define as \\textit{directed $1$-down random walk} follows a two-stage procedure at every step. The walk first selects one of the two lower-adjacent nodes with equal probability $0.5$ each, then moves towards the neighboring edges connected with the selected node with equal probabilities. If there are no other edges connected to the selected node, the walk returns to the original edge. On the other hand, the second type, which we denote as \\textit{undirected $1$-down random walk}, chooses the two nodes $u,v$ with probabilities proportional to their degrees minus one (since we want to exclude the case of returning to $e$ itself). Consequently, the walk transits to all $1$-down neighbors of the source edge with equal probabilities.", "candidate_comments": ["The two variations of edge-level random walk via down neighbors differ in whether two lower adjacent nodes of the edge have the same status. Concretely, the first type of edge-level random walk based on $\\mathbf P_{down}$, which we define as \\textit{directed $1$-down random walk} follows a two-stage procedure in every step. Starting from edge $e=(u,v)$, the walk first selects one of the two lower-adjacent nodes $u, v$ with equal probability $p(u)=p(v)=\\frac{1}{2}$. Then from the selected node, say $u$, if there are other $1$-down neighbors connect with $u$ except $e$, the walk targets at these neighboring edges with equal probabilities $\\frac{1}{d(u)-1}$ unless there are no other edges connected with the selected node. In the latter case, the walk return to the original edge $e$ with probability $1$. On the other hand, the second type which we denote as \\textit{undirected $1$-down random walk} choose the two nodes $u,v$ with probabilities proportional to their degrees minus one (since we want to exclude the case of returning to $e$ itself), i.e. $p(u):p(v)=(d(u)-1):(d(v)-1)$. Then the walk targets at the edges except $e$ that are connected with the selected nodes with equal probabilities. Since the graph is connected, the case $d(u)=d(v)=1$ will not occur as long as there are more than one edge."]}
{"annotations": [{"id": 86868399, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86901761, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86903713, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19285-66c30bd111f2a070.jsonl", "id_parag": "2310.19285-66c30bd111f2a070_5", "final_parag": "\\subsection{Sign and basis invariant edge-level positional encoding}", "candidate_comments": ["\\cai{We will apply Hodge theory and design edge-level positional encoding for graphs other than simplicials only.}"]}
{"annotations": [{"id": 86984161, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986052, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86986147, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19285-66c30bd111f2a070.jsonl", "id_parag": "2310.19285-66c30bd111f2a070_6", "final_parag": "We can show that the nonzero eigenvalues of $\\mathbf L_{1,down}$ are the same as $\\mathbf L_{0,up}$ and hence $\\mathbf L_0$. This implies that if there are no $2$-simplicials (triangles), Hodge $1$-Laplacian has the same nonzero eigenvalues as Hodge $0$-Laplacian. However, the corresponding eigenvectors still provide different information about the nodes and edges, respectively.", "candidate_comments": ["\\cref{eigen_nonzero_up_plus_down} points out that the non-zero eigenvalues of $L_1$ contain and only contain all non-zero eigenvalues of $L_{1,up}$ and $L_{1, down}$. Further \\cref{updown_nonzero_equal} gives that non-zero eigenvalues of $L_{1,down}$ is the same as $L_{0,up}$ and hence $L_0$ since $L_0=L_{0,up}$. This implies that if there are no $2$-simplicials (triangles), Hodge $1$-Laplacian has the same non-zero eigenvalues as Hodge $0$-Laplacian. However, the corresponding eigenvectors still provide different information of nodes and edges respectively. "]}
{"annotations": [{"id": 86962589, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86973531, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973856, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19285-66c30bd111f2a070.jsonl", "id_parag": "2310.19285-66c30bd111f2a070_7", "final_parag": "Similarly to the edge-level random walk in the lifted space, the transition matrix $\\hat{\\mathbf P_k}$ describes that each step of $k$-th order random walk move towards either $k$-down neighbors or $k$-up neighbors. When going through the upper adjacent $k+1$ faces, the walk uniformly transits to an upper adjacent $k$-simplex with different orientation relative to the shared $k+1$ face, unless it has no upper adjacent faces. If the step is taken towards a lower-adjacent $k-1$ face, the walk transits along or against the original direction to one of its $k$-down neighbors.", "candidate_comments": ["The full transition matrix $\\hat{\\mathbf P_k}=\\frac{1}{2}\\mathbf P_{k,down}+\\frac{1}{2}\\mathbf P_{k,up}$ describes that each step of $k$-th order random walk move towards either $k$-down neighbors or $k$-up neighbors with probability $0.5$ each. When going via upper adjacent $k+1$ faces, the walk uniformly transits to an upper adjacent $k$-simplex with different orientation relative to the shared $k+1$ face, unless it has no upper adjacent faces. If the step is taken towards lower adjacent $k-1$ face, the walk transit along or against the original direction to the lower $k-1$ faces connected with it (its boundaries) with each probability $0.5$, then further transmit to the target $k$-down neighbors with probabilities proportional to their upper degrees or the weights."]}
{"annotations": [{"id": 86904715, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86965140, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86965292, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19285-66c30bd111f2a070.jsonl", "id_parag": "2310.19285-66c30bd111f2a070_8", "final_parag": "There is an interesting and well-known connection between mean commute time and resistance between nodes $r_{xy}$, see \\citep{resistancecommute, RethinkingBiconnectivityResitance}.", "candidate_comments": ["\\cai{The mean commute time and resistance distance have a close connection, see electric networks and \\citep{RethinkingBiconnectivityResitance}}"]}
{"annotations": [{"id": 86981761, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986600, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86986567, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19285-66c30bd111f2a070.jsonl", "id_parag": "2310.19285-66c30bd111f2a070_9", "final_parag": "In addition to making use of eigenvalues, we can build more powerful GNNs based on the spectra of Hodge $k$-Laplacians. We address that we can universally approximate permutation equivariant and basis invariant functions (moreover, $k$-form) defined on $k$-simplicial complexes if we use Expressive-BasisNets \\citep{SignandBasis}. Universality is guaranteed by the following decomposition theorem.", "candidate_comments": ["\\cai{Proof for the more general version: universally approximate permutation equivariant and basis invariant functions if we use the eigenvectors properly, see \\citep{SignandBasis}; the eigenvalues of $\\Delta_k$ (normalized Hodge Laplacian in \\citep{SpectraOfCombinatorialLaplace}).}"]}
{"annotations": [{"id": 86871002, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}, {"id": 86904832, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86972223, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}], "agreement": 55.55555555555556, "source_file": "2310.19288-9ff1fd0d638c0e77.jsonl", "id_parag": "2310.19288-9ff1fd0d638c0e77_1", "final_parag": "\\thanks{This work was supported in part by the National Natural Science Foundation of China under Grant 42230108 and 61971319. (\\emph{Corresponding author: Qiangqiang Yuan}.)} \\thanks{Yi Xiao, Qiangqiang Yuan, Jiang He, and Xianyu Jin are with the School of Geodesy and Geomatics, Wuhan University, Wuhan 430000, China (e-mail: xiao\\_yi@whu.edu.cn; yqiang86@gmail.com; jiang\\_he@whu.edu.cn; jin\\_xy@whu.edu.cn).}", "candidate_comments": ["\\thanks{Xin Su is with the School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, China (e-mail: xinsu.rs@whu.edu.cn).} \\thanks{Qiangqiang Yuan is with the School of Geodesy and Geomatics, Wuhan University (e-mail: yqiang86@gmail.com).}", "\\thanks{Xianyu Jin is with the School of Geodesy and Geomatics, Wuhan University (e-mail: jin\\_xy@whu.edu.cn).} \\thanks{Jiang He is with the School of Geodesy and Geomatics, Wuhan University (e-mail: jiang\\_he@whu.edu.cn).}"]}
{"annotations": [{"id": 86723402, "completed_by": {"id": 70661}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86727354, "completed_by": {"id": 140156}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86825986, "completed_by": {"id": 126844}, "result": ["Différent", "Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19288-9ff1fd0d638c0e77.jsonl", "id_parag": "2310.19288-9ff1fd0d638c0e77_2", "final_parag": "\\thanks{Kui Jiang is with the School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150000, China (e-mail: kuijiang\\_1994@163.com).}", "candidate_comments": ["\\thanks{Xin Su is with the School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, China (e-mail: xinsu.rs@whu.edu.cn).} \\thanks{Qiangqiang Yuan is with the School of Geodesy and Geomatics, Wuhan University (e-mail: yqiang86@gmail.com).}", "\\thanks{Xianyu Jin is with the School of Geodesy and Geomatics, Wuhan University (e-mail: jin\\_xy@whu.edu.cn).} \\thanks{Jiang He is with the School of Geodesy and Geomatics, Wuhan University (e-mail: jiang\\_he@whu.edu.cn).}", "\\thanks{Chia-Wen Lin is with the Department of Electrical Engineering and the Institute of Communications Engineering, National Tsing Hua University (e-mail: cwlin@ee.nthu.edu.tw).}}"]}
{"annotations": [{"id": 86972965, "completed_by": {"id": 62306}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86973619, "completed_by": {"id": 140156}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86898976, "completed_by": {"id": 62471}, "result": ["Différent", "Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19288-9ff1fd0d638c0e77.jsonl", "id_parag": "2310.19288-9ff1fd0d638c0e77_3", "final_parag": "\\thanks{Liangpei Zhang is with the State Key Laboratory of Information Engineering in Surveying, Mapping, and Remote Sensing, Wuhan University, Wuhan 430000, China (e-mail: zlp62@whu.edu.cn).}}", "candidate_comments": ["\\thanks{Xin Su is with the School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, China (e-mail: xinsu.rs@whu.edu.cn).} \\thanks{Qiangqiang Yuan is with the School of Geodesy and Geomatics, Wuhan University (e-mail: yqiang86@gmail.com).}", "\\thanks{Xianyu Jin is with the School of Geodesy and Geomatics, Wuhan University (e-mail: jin\\_xy@whu.edu.cn).} \\thanks{Jiang He is with the School of Geodesy and Geomatics, Wuhan University (e-mail: jiang\\_he@whu.edu.cn).}", "\\thanks{Chia-Wen Lin is with the Department of Electrical Engineering and the Institute of Communications Engineering, National Tsing Hua University (e-mail: cwlin@ee.nthu.edu.tw).}}"]}
{"annotations": [{"id": 86962985, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86983720, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86983838, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19297-152045bc1a8a5cf3.jsonl", "id_parag": "2310.19297-152045bc1a8a5cf3_1", "final_parag": "unlike $\\texttt{Gender}$ and $\\texttt{Blackhair}$, who have relatively negligible skew, $\\texttt{Young}$ and $\\texttt{Attractive}$ observes a significantly larger error at $\\bpstar=[0.5,0.5]$.", "candidate_comments": ["we observe that $skew_\\balpha$ has influence on the errors observed in the lower biases ($\\pstar=[0.5,0.5]$). "]}
{"annotations": [{"id": 86981791, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86985495, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86985653, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19297-19d90053e8d4bf5a.jsonl", "id_parag": "2310.19297-19d90053e8d4bf5a_1", "final_parag": "\\section{Details on Applying CLIP as a SA Classifier}", "candidate_comments": ["\\textcolor{red}{To Include discussion on how we utilized CLIP as a attribute Classifier[{\\bf Addressed}]}"]}
{"annotations": [{"id": 86904822, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973238, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86973456, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19297-19d90053e8d4bf5a.jsonl", "id_parag": "2310.19297-19d90053e8d4bf5a_2", "final_parag": "{\\bf CLIP as a Sensitive Attribute classifier.} To utilize CLIP as a sensitive attribute classifier (with the VIT-B/32 architecture), we follow the best practices suggested by Radford \\etal \\cite{radfordLearningTransferableVisual2021}. Here, we first input two different prompts, describing the respective classes, to the CLIP text-encoder, as seen in Tab. \\ref{tab:CLIPPrompts}. As suggested by Radford \\etal we utilize the prompt starting with \"A photo of a\" \\ie a scene description, followed by our sensitive attribute's classes \\eg female/male. \nNext, we also encode the generated images with the CLIP image encoder.", "candidate_comments": ["} between the encoded generated image and the encoded text to assign the image its respective class. "]}
{"annotations": [{"id": 86829713, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86856436, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86898607, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19297-19d90053e8d4bf5a.jsonl", "id_parag": "2310.19297-19d90053e8d4bf5a_3", "final_parag": "Finally, for each encoded generated image and the two encoded text-prompt, we take the cosine similarities followed by the $\\argmax$. The $\\argmax$ output provides us with the respective hard label of the generated image.", "candidate_comments": ["} between the encoded generated image and the encoded text to assign the image its respective class. "]}
{"annotations": [{"id": 87255006, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87259320, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87267688, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19297-19d90053e8d4bf5a.jsonl", "id_parag": "2310.19297-19d90053e8d4bf5a_4", "final_parag": "{\\bf Generated Image pre-processing.} We remark that as the stable diffusion model produces a mixture of colored and greyscale images, for a fair comparison, we transform all images from RGB to greyscale before feeding into CLIP for classification.", "candidate_comments": ["Here, we utilise the prompts \"a photo of a female\" and \"a photo of a male\" to differentiate between the different classes for \\texttt{Gender}.  Then as the stable diffusion model produces a mixture of colored and greyscale images, for a fair comparison, we transform all images from rgb to greyscale before being input in CLIP for classification.  See \\ref{fig:moreIllusratioofSD} for some illustration. \nFinally, to assign the respective label, we follow radford \\etal and assign the label of the encoded prompt with the smallest cosine similarity to the encoded image. "]}
{"annotations": [{"id": 86963052, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86982428, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86982528, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19297-2ccc0505c1d71d96.jsonl", "id_parag": "2310.19297-2ccc0505c1d71d96_1", "final_parag": "the data bias, while CLEAM remains consistently low. \nSee Sec. \n\\ref{subsec:pseudoGMoreLA} and \\ref{subsec:pseudoGMoreClassifiers} for analysis with more attributes and classifiers.", "candidate_comments": ["the same ResNet-18 as per Sec. \\ref{subsec:biasmitigation}, to evaluate our pseudo-generator.  See Supp. \\ref{subsec:pseudoGMoreLA} and \\ref{subsec:pseudoGMoreClassifiers} for more analysis. "]}
{"annotations": [{"id": 86981982, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86983556, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86984203, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19297-30bedf970eb2b3b6.jsonl", "id_parag": "2310.19297-30bedf970eb2b3b6_1", "final_parag": "These samples are then hand labeled \\wrt the sensitive attributes. More specifically, \\texttt{Gender} and \\texttt{BlackHair} for both the GANs and \\texttt{Gender} for the SDM.", "candidate_comments": ["} These samples are then hand labeled \\wrt the sensitive attribute \\texttt{Gender} and \\texttt{Blackhair}, which is perceived to be consistent with human perception. "]}
{"annotations": [{"id": 86724067, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86727197, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86728692, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19297-30bedf970eb2b3b6.jsonl", "id_parag": "2310.19297-30bedf970eb2b3b6_2", "final_parag": "We remark that the discarded samples consist only a small portion of the generated samples \\ie 3\\% of the GANs, and 1\\% of the SDM. Upon further evaluation, we found that the sensitive attribute classifiers appear to uniformly assign these (rejected) ambiguous samples a random class with low confidence. As a result, we can assume that the impact of disregarding these samples was insignificant to CLEAM's evaluation.", "candidate_comments": ["} \\textcolor{blue}{Furthermore, in addition to the fact that these samples make a small portion of generated samples, a further evaluation found that the classifiers appears to uniformly assign these (rejected) ambiguous samples a random class with low confidence. As a result, we can assume that the impact of disregarding these samples was insignificant to CLEAM's evaluation. }"]}
{"annotations": [{"id": 86723821, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86724254, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86725151, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19297-5f6a58bcca89e322.jsonl", "id_parag": "2310.19297-5f6a58bcca89e322_1", "final_parag": "fine-tune the VGG-16 model on the CelebA dataset with the respective sensitive attribute. Then we removed the last fully connected layer of the classifier model, and utilise the 4096 feature vector for the diversity measurement, as per \\cite{keswaniAuditingDiversityUsing2021}. Our results in Fig. \\ref{fig:VGG16CelebA} demonstrate significant improvement on both $\\texttt{Gender}$ and $\\texttt{BlackHair}$, based on the new improved VGG-16 model implementation. This thereby verifies our intuition that there exists a mismatch of domains in the VGG-16 pretrained on ImageNet when utilized with CelebA samples.", "candidate_comments": ["As per the senstive attribute \\texttt{BlackHair}, we first re-trained the VGG-16 model on the respective SA and measured their accuracies on a validation dataset.  We noted an average accuracy of $78.44\\%$ and $84.41\\%$ for $\\texttt{Young}$ and $\\texttt{Attractive}$. Next, we removed the last layer of the classifier and utilised the 4096 feature vector for the implementation of diversity, as per \\cite{keswaniAuditingDiversityUsing2021}. Unlike the original classifier (with the last fully connected layer) the same models were unable to accurately measure the respective $\\pstar_0$ (biases) with the diversity algorithm, as seen in Fig. \\ref{fig:VGG16AttractiveYoung}. From this, we observe that even our fined tuned VGG-16 feature extractor did not improve the Diversity framework on these harder SA."]}
{"annotations": [{"id": 86737127, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86901481, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86901598, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19297-5f6a58bcca89e322.jsonl", "id_parag": "2310.19297-5f6a58bcca89e322_2", "final_parag": "However, upon further experimentation, we recognize certain limitations still exist in the Diversity measure when used on more ambiguous and harder sensitive attribute \\eg $\\texttt{Young}$ and $\\texttt{Attractive}$. Similar to before, we fine-tuned the sensitive attribute classifier (feature extractor) which achieved accuracies of $78.44\\%$ and $84.41\\%$ for $\\texttt{Young}$ and $\\texttt{Attractive}$, respectively. However even with this re-implementation, the diversity persistent to perform poorly, as seen in Fig. \\ref{fig:VGG16AttractiveYoung}.", "candidate_comments": ["As per the senstive attribute \\texttt{BlackHair}, we first re-trained the VGG-16 model on the respective SA and measured their accuracies on a validation dataset.  We noted an average accuracy of $78.44\\%$ and $84.41\\%$ for $\\texttt{Young}$ and $\\texttt{Attractive}$. Next, we removed the last layer of the classifier and utilised the 4096 feature vector for the implementation of diversity, as per \\cite{keswaniAuditingDiversityUsing2021}. Unlike the original classifier (with the last fully connected layer) the same models were unable to accurately measure the respective $\\pstar_0$ (biases) with the diversity algorithm, as seen in Fig. \\ref{fig:VGG16AttractiveYoung}. From this, we observe that even our fined tuned VGG-16 feature extractor did not improve the Diversity framework on these harder SA."]}
{"annotations": [{"id": 86972265, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86979424, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86979904, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19297-a54117f4ad9b2649.jsonl", "id_parag": "2310.19297-a54117f4ad9b2649_1", "final_parag": "[EQUATION] where $c_{i|j}$ denotes the event of assigning label $i$ to a sample with GT label $j$. \nThen, we mentioned that the count for each ouptut can be modeled as a multinomial distribution, $\\mathbf{N}_{\\bc}\\sim Multi(n,\\bp)$. Note that $\\mathbf{N}_\\bc = [N_{c_{0|0}},N_{c_{1|0}},N_{c_{1|1}},N_{c_{0|1}}]^T $ is the random vector of counts for individual outputs of $\\bc$.", "candidate_comments": ["models the {\\em joint probability distribution} of these outputs, \\ie $ \\mathbf{N}_\\bc= \\mathbb{P}(N_{c_{0|0}},N_{c_{1|0}},N_{c_{1|1}},N_{c_{0|1}})$ where, "]}
{"annotations": [{"id": 86739474, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86768629, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86827420, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19297-a54117f4ad9b2649.jsonl", "id_parag": "2310.19297-a54117f4ad9b2649_2", "final_parag": "[EQUATION] and therefore: [EQUATION] [EQUATION] With this, we note that the {\\em marginal distribution} of this multivariate Gaussian distribution gives us a univariate (one-dimensional) Gaussian distribution for the count of each output in $\\bc^T$. For example, the distribution of the count for event $c_{0|0}$, denoted by $N_{c_{0|0}}$, can be modeled as $N_{c_{0|0}} \\sim \\mathcal{N}(\\boldsymbol{\\mu}_1, \\boldsymbol{\\Sigma}_{11})$. \nThen, we find the total rate of data points labeled as class $i$ when labeling $n$ generated images using the normalized sum of the related random variables \\ie $\\hat{p}_i=\\frac{1}{n}\\sum_j N_{{c}_{i|j}}$. More specifically:", "candidate_comments": ["With that, we solve for the total rate of data points labelled as class $i$ when labelling $n$ generated samples using the normalized sum of the related random variable \\ie  We then normalise $\\textbf{N}_{\\phat_i}$, to get $\\phat_i=\\frac{1}{n}\\textbf{N}_{\\phat_i},\\;i\\in\\{0,1\\}$ via the marginal distribution of $\\mathbf{N}_{\\phat_0}$ and $\\mathbf{N}_{\\phat_1}$ where, "]}
{"annotations": [{"id": 86962725, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86985221, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86989703, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19297-a54117f4ad9b2649.jsonl", "id_parag": "2310.19297-a54117f4ad9b2649_3", "final_parag": "{\\bf Remark:} In Sec 4.1 of the main paper, considering the probability tree diagram in Fig.1(b) (main paper), we proposed a distribution for the possible events of classification ($c_{i|j}$), and used it to compute distribution of each event, and finally the distribution of the output of the sensitive attribute classifier ($\\phat_0$, and $\\phat_1$). Here, we provide more information on the necessary assumptions and the expanded forms of the equations. \nIn the following Sec. \\ref{apx:proposedmethod}, we will similarly provide more information on proposed CLEAM, presented in Sec. 4.2 of the main paper, which utilizes this statistical model to mitigate the sensitive attribute classifier's error.", "candidate_comments": ["From this, given that $s$ is large, we assume that the sample mean is a good approximation for the population mean (modeled in Sec. \\ref{subsec:expandedFormCLEAM} ) and equate them  \\textcolor{red}{\\bf we have to provide more information as to how we got to the RHS of the equation}:"]}
{"annotations": [{"id": 86898924, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86901361, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86907388, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19297-a54117f4ad9b2649.jsonl", "id_parag": "2310.19297-a54117f4ad9b2649_4", "final_parag": "{\\bf Point Estimate of CLEAM.} From this, given that $s$ is sufficiently large, we utilize the sample mean as the maximum likelihood approximate of the population mean. As the population mean was modeled in Sec. \\ref{subsec:expandedFormCLEAM}, we can equate the sample mean to the expanded theoretical model:", "candidate_comments": ["From this, given that $s$ is large, we assume that the sample mean is a good approximation for the population mean (modeled in Sec. \\ref{subsec:expandedFormCLEAM} ) and equate them  \\textcolor{red}{\\bf we have to provide more information as to how we got to the RHS of the equation}:"]}
{"annotations": [{"id": 86960881, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973415, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974129, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19297-c41bd771c56fa17e.jsonl", "id_parag": "2310.19297-c41bd771c56fa17e_1", "final_parag": "\\textbf{Fairness in Generative Models.} Fairness in machine learning is mostly studied for discriminative learning, where usually the objective is to handle a classification task independent of a sensitive attribute in the input data, \\eg making a hiring decision independent of the applicant \\texttt{Gender}. However, the definition of fairness is quite different for generative learning, where it is considered as equal representation/generation probability \\wrt a sensitive attribute.", "candidate_comments": ["for the training of a Fair-GAN model. \nIn this algorithm, a reference fair dataset \\wrt the sensitive attribute is used during training, while still exposing the model to the larger biased dataset, but down-weighting it. This allows the generator to output high-quality samples while encouraging fairness \\wrt the sensitive attribute."]}
{"annotations": [{"id": 86962509, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86983235, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86983337, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19297-c41bd771c56fa17e.jsonl", "id_parag": "2310.19297-c41bd771c56fa17e_2", "final_parag": "\\textbf{Biases in Text-Image generation.} Some literature have attempted to look into the biases in text-to-image generators \\cite{bianchiEasilyAccessibleTexttoImage2022}. Specifically, Bianchi \\etal study existing biases in occupations-based prompts for popular text-to-image generators \\eg stable diffusion models. They found the biases to exasperate existing occupation stereotypes, \\eg nurses being over-represented as non-Caucasian females. \nTo measure these biases, \\cite{bianchiEasilyAccessibleTexttoImage2022} has a simple approach utilizing a pre-trained feature extractor to assign the sensitive attribute labels to a small batch of generated images (100 samples).", "candidate_comments": ["based on their similarity to a labelled reference dataset. \nTo measure these biases the work compares the latent-embeddings, utilizing a pre-trained feature extractor, of a small batch of generated image (100 samples) against a reference dataset \\wrt a sensitive attributes and assign the respective labels based on a similarly metric. \n} \\textcolor{blue}{ 1) Differences in the measurements method to our proposed method"]}
{"annotations": [{"id": 86981613, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86983398, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86983522, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19297-c41bd771c56fa17e.jsonl", "id_parag": "2310.19297-c41bd771c56fa17e_3", "final_parag": "We remark that this approach is similar to Diversity, a method which we found to also demonstrate significant errors due to the lack of consideration for the classifier's error.", "candidate_comments": ["We remark that, in our work, we look into the accuracy of similar classification methods (Diversity). Here, we found that these bias measurement methods also demonstrate significant errors due to the lack of consideration for the classifiers error.  CLEAM is then propose to improve upon these erroneous measurements. \n2) Differences in the they studied application and ours Then in "]}
{"annotations": [{"id": 86725184, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86727437, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86728962, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19297-c41bd771c56fa17e.jsonl", "id_parag": "2310.19297-c41bd771c56fa17e_4", "final_parag": "Furthermore, we emphasize the difference between our study and Bianchi \\etal. Specifically, in our application of CLEAM (Sec. 6 of the main manuscript), we examine the impact of using prompts with indefinite pronouns/nouns that are synonymous to each another. \nOur objective, unlike Bianchi \\etal's work, is to investigate the influence of subtle changes in the prompts on bias, which is studied on a large dataset ($\\approx 2k$ samples). Our results are the first to demonstrate that even subtle changes to the prompt (which are semantically unchanged), could result in drastically different biases.", "candidate_comments": ["In the application section (Sec. 6 of the main manuscript), we examine the impact of using prompts with indefinite pronouns/nouns that are synonymous with each other. Our objective, in contrast to Bianchi et al.'s work, is to investigate the influence of various synonymous indefinite (neutral) prompts on bias. Surprisingly, our findings demonstrate that these different synonymous prompts exhibit varying degrees of bias. This outcome contradicts our initial expectations, as we anticipated that the semantically similar prompts would produce similar biases, and since they were neutral, the biases would be negligible."]}
{"annotations": [{"id": 87255090, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87259698, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87267796, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19297-c41bd771c56fa17e.jsonl", "id_parag": "2310.19297-c41bd771c56fa17e_5", "final_parag": "More specifically, conventional classifier calibration methods usually work on soft labels (probabilities). Note that in our framework, the $argMax$ is applied to the output probabilities to determine the hard label. Therefore, in our application that deals with hard labels, regular classification techniques are less effective. \nTo investigate this, we conduct a few calibration experiment by applying some popular classifier calibration techniques; temperature scaling(T-scaling) \\cite{guoCalibrationModernNeural2017}, Isotonic Regression\\cite{nyberg2021reliably} and Platt Scaling\\cite{platt1999probabilistic} on a pre-trained ResNet-18\\cite{heDeepResidualLearning2016} senstive attribute classifier. In Fig. \\ref{fig:Calibration_curve}, we see that T-scaling is the most effective in correcting the calibration curve to the ideal Ref line. Note that, this Ref line indicates that the classifier is perfectly calibrated \\wrt the soft labels.", "candidate_comments": ["However, as shown in Fig. \\ref{fig:Calibration_Error} T-scaling shows negligible improvement in normalized error, demonstraing that traditional calibration technique may not have a direct correlation to hard label calibration, which CLEAM aims to address."]}
{"annotations": [{"id": 86962079, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86975176, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86988323, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19297-c41bd771c56fa17e.jsonl", "id_parag": "2310.19297-c41bd771c56fa17e_6", "final_parag": "\\textbf{In our results,} seen in Fig. \\ref{fig:Calibration_Error}, we observe that these traditional calibration methods are less effective in correcting the sensitive attribute distribution error. In fact, methods like Platt scaling worsen the error, and T-scaling ---which is shown in \\cite{guoCalibrationModernNeural2017} and our experiment to be one of the most effective traditional calibration methods--- does not change class predictions (hard labels), but merely perturb the soft labels. This demonstrates that traditional calibration technique are not direct correlation to hard label calibration, which CLEAM aims to address.", "candidate_comments": ["However, as shown in Fig. \\ref{fig:Calibration_Error} T-scaling shows negligible improvement in normalized error, demonstraing that traditional calibration technique may not have a direct correlation to hard label calibration, which CLEAM aims to address."]}
{"annotations": [{"id": 86730732, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86724635, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86762117, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19297-c5eddd17607c84cc.jsonl", "id_parag": "2310.19297-c5eddd17607c84cc_1", "final_parag": "To do this, we first obtain $s=30$ values of $\\bphat$ from framework shown in Fig. 1 of the main paper, and use StyleGAN2 \\cite{karrasStyleBasedGeneratorArchitecture2019} and StyleSwin \\cite{zhang2021styleswin} as the generative model. \nThen using ResNet-18 with known $\\balpha$ and GAN's GT $\\bpstar$, as discussed in Sec. 4.1 of the main paper, we form the theoretical model's Gaussian distribution, $\\mathcal{N}(\\tilde{\\mu}_{\\phat_0},\\tilde{\\sigma}^2_{\\phat_0})$.", "candidate_comments": ["-------------Setup/objective that we agreed upon-------------------------- Recall that we showed in Sec. \\ref{sec:sampleBasedEstimate} that the sample-based estimate is a good approximation of the theoretical model. Hence, we use the sample-based estimates in Eqn. 5 and Eqn. 6 of the main paper to approximated mean and variances of the theoretical model's Gaussian distribution, $\\mathcal{N}(\\Ddot{\\mu}_{\\phat_0},\\Ddot{\\sigma}^2_{\\phat_0})$."]}
{"annotations": [{"id": 86952384, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86959053, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86961828, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19297-c5eddd17607c84cc.jsonl", "id_parag": "2310.19297-c5eddd17607c84cc_2", "final_parag": "As seen in Fig. \\ref{fig:QQplot}, the $\\bphat$ samples from GAN correlate tightly with the standardised line (in red), a line indicating a perfect correlation between theoretical and sample quantiles. This analysis supports our claim that the $\\bphat$ samples from a real generator (GAN) follow the distribution estimated by the proposed model.", "candidate_comments": ["the  $\\phat$ samples from the  real GAN generators in \\cite{choiFairGenerativeModeling2020a} \\textcolor{red}{(Can we say this a bit general? e.g.: 'the $\\phat$ samples from a real generator (GAN) follow the distribution estimated by the proposed model.')} follow a Gaussian distribution."]}
{"annotations": [{"id": 86731793, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86731209, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}, {"id": 86732458, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19297-c7854825ef6e28bb.jsonl", "id_parag": "2310.19297-c7854825ef6e28bb_1", "final_parag": "Then similar to our previous discussion, we would also refrain from ambiguous sensitive attributes, as this may result in a mismatch between the proxy validation dataset and the pre-trained sensitive attribute classifier.", "candidate_comments": ["Ambigious sensitive attributes may mean different things between the validation dataset and the pre-trained classifier", "we would recommend to refrain from measuring ambiguous sensitive attributes, as discussed earlier, as the proxy validation dataset and the generated images  }"]}
{"annotations": [{"id": 86734768, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86901646, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86902196, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19297-cf39802279ee1ee0.jsonl", "id_parag": "2310.19297-cf39802279ee1ee0_1", "final_parag": "\\textbf{In our results,} seen in Fig. \\ref{fig:Calibration_Error}, we observe that these traditional calibration methods are less effective in correcting the LA distribution error. In fact, methods like Platt scaling worsen the error, and T scaling ---which is shown in \\cite{guoCalibrationModernNeural2017} and our experiment to be one of the most effective traditional calibration methods--- does not change class predictions (hard labels), but merely perturb the soft labels. This demonstrates that traditional calibration technique may not have a direct correlation to hard label calibration, which CLEAM aims to address.", "candidate_comments": ["However, as shown in Fig. \\ref{fig:Calibration_Error} T-scaling shows negligible improvement in normalized error, demonstraing that traditional calibration technique may not have a direct correlation to hard label calibration, which CLEAM aims to address."]}
{"annotations": [{"id": 86983312, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86982314, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86983197, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19297-cf39802279ee1ee0.jsonl", "id_parag": "2310.19297-cf39802279ee1ee0_2", "final_parag": "As seen in Fig. \\ref{fig:QQplot}, the $\\phat$ samples from GAN correlate tightly with the standardised line (in red), a line indicating a perfect correlation between theoretical and sample quantiles. This analysis supports our claim that the $\\phat$ samples from a real generator (GAN) follow the distribution estimated by the proposed model.", "candidate_comments": ["the  $\\phat$ samples from the  real GAN generators in \\cite{choiFairGenerativeModeling2020a} \\textcolor{red}{(Can we say this a bit general? e.g.: 'the $\\phat$ samples from a real generator (GAN) follow the distribution estimated by the proposed model.')} follow a Gaussian distribution."]}
{"annotations": [{"id": 86722904, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86732380, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86827925, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19297-cf39802279ee1ee0.jsonl", "id_parag": "2310.19297-cf39802279ee1ee0_3", "final_parag": "\\textbf{\\em In this setup,} we repeat the same experiment with the pseudo-generator, but instead on LA $\\texttt{Young}$ and $\\texttt{Attractive}$ from the CelebA dataset. Given that both LA classifiers trained on either $\\texttt{Young}$ or $\\texttt{Attractive}$ have similar average accuracy, $\\alpha_{Avg}=\\frac{\\alpha_0+\\alpha_1}{2}$ of $0.801$ and $0.794$ but different $skew_\\balpha$ of 0.103 and 0.027, we are able to investigate the effects of $skew_\\balpha$ on both CLEAM and Baseline. We did not include Diversity in this study, due to its poor performance on harder LA, as discussed in \\ref{sec:DiversitySetup}.", "candidate_comments": ["in Tab. \\ref{tab:peYoungAttractive}, \\ref{tab:ieYoungAttractive}, who's $\\balpha$ are skewed at different degrees, although having similar $\\alpha_{Avg}=\\frac{\\alpha_0+\\alpha_1}{2}$ of $0.801$ and $0.794$ respectively. The skew in $\\balpha$ is defined as the difference between $\\alpha_0$ and $\\alpha_1$ \\ie skew=$|\\alpha_0-\\alpha_1|$."]}
{"annotations": [{"id": 86725620, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86725799, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86727190, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19297-d0ad67c87bdb7876.jsonl", "id_parag": "2310.19297-d0ad67c87bdb7876_1", "final_parag": "In our proposed CLEAM, we use $\\balpha$ pre-measured on the validation dataset, denoted by $\\balpha_{val}$. In this section, we show that $\\balpha_{val}$ is a good approximate of the $\\balpha$ when measured on the generated data, denoted by $\\balpha_{gen}$.", "candidate_comments": ["Note that $\\balpha_{gen}$ is unknown in practice, therefore $\\balpha_{val}$ is used instead.  As an accurate $\\balpha$ is a crucial component to CLEAM, it is important to understand if our pre-evaluated $\\balpha$ on the validation dataset, denoted by $\\balpha_{val}$, is a good approximate of the $\\balpha$ when measured on the generated data, denoted by $\\balpha_{gen}$. \nNote that this is important as generally,"]}
{"annotations": [{"id": 86961714, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86983382, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86983459, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19297-d0ad67c87bdb7876.jsonl", "id_parag": "2310.19297-d0ad67c87bdb7876_2", "final_parag": "Note that $\\balpha_{gen}$ is not available in practice, therefore $\\balpha_{val}$ is used as approximation during CLEAM measurement. We further remark that our purpose of this experiment is only done to validate $\\balpha_{val}$ as a good approximation for $\\balpha_{gen}$ and is not necessary in the actual deployment of CLEAM.", "candidate_comments": ["Note that $\\balpha_{gen}$ is unknown in practice, therefore $\\balpha_{val}$ is used instead.  As an accurate $\\balpha$ is a crucial component to CLEAM, it is important to understand if our pre-evaluated $\\balpha$ on the validation dataset, denoted by $\\balpha_{val}$, is a good approximate of the $\\balpha$ when measured on the generated data, denoted by $\\balpha_{gen}$. \nNote that this is important as generally,"]}
{"annotations": [{"id": 86961105, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972802, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86972960, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19297-d0ad67c87bdb7876.jsonl", "id_parag": "2310.19297-d0ad67c87bdb7876_3", "final_parag": "As shown in Tab. \\ref{tab:approxalpha} our approximated $\\balpha_{val}$ (measured on CelebA-HQ), although not perfect, is a close approximation of $\\balpha_{gen}$, thereby making it appropriate to be utilized with CLEAM.", "candidate_comments": ["Similarly to attain the approximated $\\balpha_{val}$, as discussed in the main manuscript when a validation dataset isn't available for SDMs, "]}
{"annotations": [{"id": 86960858, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86982062, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86982081, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86988902, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19297-d8d5c9fa812ec1e9.jsonl", "id_parag": "2310.19297-d8d5c9fa812ec1e9_1", "final_parag": "Then, when analyzing bias mitigation techniques using IE of CLEAM (as per Tab. \\ref{tab:biasMitigation_IE}), since the IE of unweighted and weighted GANs do not overlap, we are provided with some statistical guarantees that the bias mitigation techniques, importance-weighting, is indeed effective.", "candidate_comments": ["This section contains the interval estimates (IE) results for the experiments in Sec. 5.2. As discussed in the main paper, the point estimate can still contain some approximation error. As a result, we suggest that in addition to the PE, the IE is better to be reported.  Our results, in the main paper (Sec 5.1 and 5.3), show that in our experiments where the GT $\\pstar$ is known CLEAM's IE, in most cases, demonstrates being able to to bound the $\\pstar$. \nThen, when analyzing bias mitigation techniques using IE of CLEAM (as per Tab. \\ref{tab:biasMitigation_IE}), since the IE of unweighted and weighted GANs do not overlap, we are provided with some statistical guarantees that the bias mitigation techniques, importance weighting, is indeed effective."]}
{"annotations": [{"id": 86721378, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86732213, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86765418, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19297-dee24728a406a7ef.jsonl", "id_parag": "2310.19297-dee24728a406a7ef_1", "final_parag": "\\section*{Acknowledgements} This research is supported by the National Research Foundation, Singapore under its AI Singapore Programmes (AISG Award No.: AISG2-TC-2022-007) and SUTD project PIE-SGP-AI-2018-01. This research work is also supported by the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (Grant No. M23L7b0021). This material is based on the research/work support in part by the Changi General Hospital and Singapore University of Technology and Design, under the HealthTech Innovation Fund (HTIF Award No. CGH-SUTD-2021-004). \nWe thank anonymous reviewers for their insightful feedback and discussion.", "candidate_comments": ["These research projects are supported by the National Research Foundation, Singapore under its AI Singapore Programs (AISG Award No.: AISG2-RP-2021-021; AISG Award No.: AISG2-TC-2022-007).], and SUTD project PIE-SGP-AI-2018-01. We thank anonymous reviewers for their insightful comments."]}
{"annotations": [{"id": 86735410, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86961722, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86959374, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19303-33858952d7d8d86d.jsonl", "id_parag": "2310.19303-33858952d7d8d86d_1", "final_parag": "In making our system work, how we implement user utterances is very important. In the case of general interactive recommendation, user speech is often implemented using a publicly available dialogue recommendation dataset. Since the data contained in a dialogue recommendation dataset is a pair of a question by the user and a response by the system, and there is no pair data of a question by the system and a response by the user, it is difficult to experiment with a public dataset. Therefore, we attempt to solve this problem by simulating users using Chat-GPT. \nWe have created a module that incorporates the user's gender, age, and occupation into the prompts and outputs the user's responses, which act as input to the assistant's utterances. This allows us to validate a dynamic dialogue system between the assistant and the controller. \nBelow are the prompts for the user simulator on the subject of apartment property recommendations as well as the controller and assistant.", "candidate_comments": ["Start by inquiring about the user's current living situation and daily routines to gain a foundational understanding of their lifestyle. This will help tailor further questions more accurately."]}
{"annotations": [{"id": 86972447, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86981687, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86981763, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19306-8189aa2a71008437.jsonl", "id_parag": "2310.19306-8189aa2a71008437_1", "final_parag": "Additional fracture tests are performed for TBGs with different lattice orientations (Supplementary Fig. 9). The interaction between cracks in neighboring graphene layers is measured by the difference in the number of kinks along the crack paths between TBG and the single layers with the same lattice orientation (Supplementary Table 1). \nThe results show that the interaction is weakened as $\\theta_\\mathrm{Z}$ deviates from $\\pm19.11^\\circ$, and becomes negligible for large deviation or two aligned cracks (e.g., $\\theta_\\mathrm{Z}=\\pm17.48^\\circ$, Supplementary Fig. 10). \nThis phenomenon may be attributed to the competition between the stress field in the opening mode (mode \\uppercase\\expandafter{\\romannumeral1}, quantified by the stress intensity factor or SIF, $\\Kone$) that preserves the crack direction, the stress field in the shear mode (mode \\uppercase\\expandafter{\\romannumeral2}, $\\Ktwo$) that deflects the crack, and the weak interlayer interaction that adds an additional driving force, $K_{\\rm vdW}$. \nAs the ratio of $\\Ktwo/\\Kone$ approaches a critical value at which crack deflection can be activated, the interlayer interaction significantly modifies the crack paths (e.g., $\\theta_\\mathrm{Z}=\\pm19.11^\\circ $). \nOtherwise, the effect of interlayer interaction becomes negligible (e.g., $\\theta_\\mathrm{Z}=\\pm27.46^\\circ $) (Fig.\\,\\ref{Fig_5}e).", "candidate_comments": ["\\pj{The evidence of $K_{\\rm vdW}$ affect the crack deflection is observed at the $\\theta_\\mathrm{Z}=\\pm9.81^\\circ $.}"]}
{"annotations": [{"id": 86722270, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86732293, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86766272, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19341-66c30bd111f2a070.jsonl", "id_parag": "2310.19341-66c30bd111f2a070_1", "final_parag": "Moreover, we notice one particular anomaly in the $\\Delta_1$ column, indicating the significantly lower $L_{test}$ loss compared to $L_{ref}$, which is interesting to further study for better understanding.", "candidate_comments": ["Moreover, the $\\Delta_1$ column indicates that models generally incur lower loss on the reference set, potentially due to the fact that those GPT-4 generated samples are comparatively simpler\\footnote{We found out that most generated problems can be solved in one or two steps.} than real GSM8K samples. However, it seems that the \"expected\" value of $\\Delta_1$ should be within the range of $[0.3, 0.4]$. One particular anomaly in the $\\Delta_1$ column is the significantly lower $L_{test}$ loss of Aquila2-34B compared to $L_{ref}$, which is interesting to further study for better understanding.."]}
{"annotations": [{"id": 86985752, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86986126, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86990043, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19341-66c30bd111f2a070.jsonl", "id_parag": "2310.19341-66c30bd111f2a070_2", "final_parag": "\\section{More Benchmark Results} We also provide results of the following benchmarks in Table \\ref{table:en_benchmark}: \\begin{itemize} \\item TriviaQA \\cite{triviaqa}: TriviaQA is a realistic text-based question answering dataset which includes 950K question-answer pairs from 662K documents collected from Wikipedia and the web. \n\\item HellaSwag \\cite{hellaswag}: HellaSWAG is a dataset that focuses on grounded commonsense inference. \n\\item Winogrande \\cite{winogrande}: WinoGrande is a dataset that focuses on commonsense reasoning. \n\\item BoolQ \\cite{boolq} BoolQ is a question answering dataset for yes/no questions. \n\\item PIQA \\cite{piqa}: PIQA is a dataset for commonsense reasoning, and was created to investigate the physical knowledge of existing models in NLP. \n\\item ARC \\cite{arc}: ARC is a dataset consisting of multiple-choice question-answering tasks that focus on commonsense reasoning.", "candidate_comments": ["\\item OBQA \\cite{obqa} OBQA is a dataset that contains questions requiring multi-step reasoning, the use of additional common and commonsense knowledge, and rich text comprehension."]}
{"annotations": [{"id": 86972487, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86984869, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86984981, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19341-66c30bd111f2a070.jsonl", "id_parag": "2310.19341-66c30bd111f2a070_3", "final_parag": "\\item RACE \\cite{race} RACE is a dataset that focuses on reading comprehension. \n\\end{itemize}", "candidate_comments": ["\\item OBQA \\cite{obqa} OBQA is a dataset that contains questions requiring multi-step reasoning, the use of additional common and commonsense knowledge, and rich text comprehension."]}
{"annotations": [{"id": 86961810, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86965634, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86970998, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19345-03cf4a9a918931f8.jsonl", "id_parag": "2310.19345-03cf4a9a918931f8_1", "final_parag": "It is composed of 200 segments that require the translation of at least one English gender-neutral word into the corresponding masculine or feminine target word(s) in German. \nThe test suite is created as an extension of MuST-SHE, a multilingual, natural benchmark built on TED talks data \\citep{bentivogli-2020-gender}, which allows for a fine-grained analysis of gender bias in MT and ST.", "candidate_comments": ["It is created as an extension of MuST-SHE,  a multilingual, natural benchmark allowing for a fine-grained analysis of gender bias in MT and ST built on TED talks data \\citep{bentivogli-2020-gender}. "]}
{"annotations": [{"id": 86963170, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86985019, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985112, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19345-03cf4a9a918931f8.jsonl", "id_parag": "2310.19345-03cf4a9a918931f8_2", "final_parag": "Once the automatic step was concluded, the pool of retrieved sentence pairs underwent a manual inspection to: \\textit{i)} remove any noise and keep only pairs containing at least one gender phenomenon; \\textit{ii)} ensure that the final (\\textit{transcript, translation}) pairs were not affected by misalignments resulting from the automatic procedure used to create MuST-C and the new TED Talks data. Also, we examined the remaining pairs to verify that those to be included in MuST-SHE featured a", "candidate_comments": ["\\bs{Once the automatic step was concluded, the pool of retrieved sentence pairs underwent a manual inspection to: \\textit{i)} remove any noise and keep only pairs containing at least one gender phenomenon; \\textit{ii)} examine the remaining pairs and include in MuST-SHE   a balanced distribution of categories, F/M forms, and speakers;\\textit{iii})  ensure that the final (\\textit{transcript, translation}) pairs were not affected by misalignments resulting from the automatic procedure used to create MuST-C and the new TED Talks data. }  ======"]}
{"annotations": [{"id": 86864972, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86899840, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86900535, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19345-03cf4a9a918931f8.jsonl", "id_parag": "2310.19345-03cf4a9a918931f8_3", "final_parag": "The manual selection of appropriate sentences and their annotation was carried out by two annotators,", "candidate_comments": ["both students with a background in Applied Linguistics and a good proficiency in the German language.\\footnote{Their work was carried out as part of an internship at FBK.}"]}
{"annotations": [{"id": 86981429, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86985200, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86985193, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19345-03cf4a9a918931f8.jsonl", "id_parag": "2310.19345-03cf4a9a918931f8_4", "final_parag": "both students proficient in the German language and with a background in Applied Linguistics.\\footnote{Their work was carried out as part of an internship at FBK.}", "candidate_comments": ["both students with a background in Applied Linguistics and a good proficiency in the German language.\\footnote{Their work was carried out as part of an internship at FBK.}"]}
{"annotations": [{"id": 86904906, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86981923, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86981945, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19345-03cf4a9a918931f8.jsonl", "id_parag": "2310.19345-03cf4a9a918931f8_5", "final_parag": "Unfortunately, we found only very few instances of double German realizations, and thus at the end of the manual revision, we remained with 162 German sentences: 21 with an inclusive source term, and 141 with a non-inclusive masculine generic in the source. All the 162 manually-curated German source sentences are included in INES, and provided with their corresponding English \\texttt{IN}/\\texttt{N-IN} term pair so as to allow for focused evaluations.", "candidate_comments": ["At the end of the manual revision, we thus obtained 162 German sentences: 21 with an inclusive source term, and 141 with a non-inclusive masculine generic in the source. All the 162 manually-curated German source sentences are included in INES, and provided with their corresponding English (\\texttt{IN}/\\texttt{N-IN} term pair so to allow for focused evaluations.  "]}
{"annotations": [{"id": 86963040, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86973115, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973466, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19345-03cf4a9a918931f8.jsonl", "id_parag": "2310.19345-03cf4a9a918931f8_6", "final_parag": "In light of the above, we also propose the \\textbf{Inclusivity Index} metric, defined as:", "candidate_comments": ["In light of the above,  we also propose the \\textbf{Inclusivity Index} metric, which is based on the proportion of \\texttt{N-IN} terms annotated in INES that are generated by % a system as follows:    "]}
{"annotations": [{"id": 86962549, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86980007, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86980959, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19345-03cf4a9a918931f8.jsonl", "id_parag": "2310.19345-03cf4a9a918931f8_7", "final_parag": "improvement of masculine/feminine gender translation into grammatical gender languages \\citep{savoldi2021gender}. \nAlong this line, different strategies have been devised to improve gender translation and mitigate masculine bias \\citep{costa-jussa-de-jorge-2020-fine,gaido-etal-2021-split, choubey2021improving, saunders-etal-2022-first}. \nTo test these methods and inspect systems' behaviour, several template-based datasets have been made available --", "candidate_comments": ["ensure that unambiguous genderedreferents (e.g. \\textit{\\textbf{he}/\\textbf{she}} \\textit{is a doctor})  are properly resolved in the target language \\citep{costa-jussa-de-jorge-2020-fine, choubey2021improving, saunders-etal-2022-first}. \nThese methods are often tested on  several synthetic template-based datasets "]}
{"annotations": [{"id": 86982760, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86984531, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86984760, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19345-03cf4a9a918931f8.jsonl", "id_parag": "2310.19345-03cf4a9a918931f8_8", "final_parag": "As far as the topic of inclusivity and neutral language translation is concerned, research in MT is quite in its infancy. A notable exception is the work by \\citet{saunders-etal-2020-neural}, who created parallel test and fine-tuning data to develop MT systems able to generate non-binary translations for English$\\rightarrow$German/Spanish. However, their target sentences are artificial -- created by replacing gendered morphemes and articles with synthetic placeholders -- thus serving only as a proof-of-concept. \\citet{piergentili-etal-2023-gender}, instead, are the first to advocate for the use of target gender-neutral rephrasings and synonyms as a viable paradigm toward more inclusive MT when gender is unknown or simply irrelevant. \\citet{cho-etal-2019measuring} and \\citet{ghosh2023chatgpt} investigate the preservation of gender-neutral pronouns for Korean/Bengali$\\rightarrow$English. Their results, however, show that current MT systems still face serious difficulties on relying on the inclusive, neutral pronoun \\textit{they} in translation. Along this line of work, INES -- to the best of our knowledge -- represents the first test suite designed to asses the use of neutral, inclusive forms beside pronouns for translating into English.", "candidate_comments": ["Such challenges are exemplified  by \\citet{saunders-etal-2020-neural}, who created parallel test and fine-tuning data to develop MT systems able to generate non-binary translations  for English$\\rightarrow$German/Spanish.  However, their  target sentences are artificial -- created by replacing gendered morphemes and articles with synthetic  placeholders -- thus serving only as a proof-of-concept.  To the best of our knowledge, \\citet{piergentili2023gender}  are the first to advocate the use of target gender-neutral rephrasings and synonyms as a viable paradigm toward more inclusive MT when gender is unknown or simply irrelevant.  Despite this call to action, no concrete steps have been taken yet to  actually  facilitate  research in this direction,  not even toward suitable benchmarks to recognize  the neutral forms occasionally generated by  current systems \\citep{savoldi-etal-2022-morphosyntactic}."]}
{"annotations": [{"id": 86986732, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86987045, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86990115, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19359-66c30bd111f2a070.jsonl", "id_parag": "2310.19359-66c30bd111f2a070_1", "final_parag": "\\textbf{Predictions at the level of instances (patches)}. \nAlthough they only use bag labels for training, the compared methods can make predictions at the level of instances, recall Section \\ref{sec:predictions}. \nThis is important to determine more precisely in which region (patch) the cancer is present. \nTable \\ref{tab:instance_results} shows the results when making predictions at patch level. We observe that VGPMIL-PR-I outperforms both baselines in all the metrics for four out of five values of $\\lambda$ (and for $\\lambda=0.1$, the baselines are only better in terms of precision). We also appreciate that VGPMIL-PR-I results are robust across different runs, obtaining low values of standard deviation. This stability is important for real-world applications, where one wants to avoid high sensitivity to random initializations. Finally, notice that, as argued in \\cite{wang2021results}, we also observe that VGPMIL-PR (slightly) outperforms VGPMIL.", "candidate_comments": ["Interestingly, the correlation introduced by the Ising model can help at this, since they contribute to keep patch-level predictions more homogeneous."]}
{"annotations": [{"id": 86734988, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86904354, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86910190, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19359-66c30bd111f2a070.jsonl", "id_parag": "2310.19359-66c30bd111f2a070_2", "final_parag": "Interestingly, the Ising model can help to avoid isolated false positive predictions on the patch-level (which lead to false positive bag predictions) by using the patch-level correlation. \nThis is reflected in Table \\ref{tab:cms}, where we observe that increasingly more negative images are predicted correctly as $\\lambda$ gets higher. \nIn contrast, notice that strong correlation damage the performance in the positive class, as they penalize the appearance of positive patches (which would break the homogeneity of the bag, where most patches do not contain cancer). Therefore, we conclude that instance correlation is beneficial when used with a low-to-medium intensity. For instance, in this application $\\lambda=0.5$ is the best performing value, and it will be the one used by default in the sequel.", "candidate_comments": ["Interestingly, the correlation introduced by the Ising model can help at this, since they contribute to keep patch-level predictions more homogeneous."]}
{"annotations": [{"id": 86962788, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86985883, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86988358, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19359-66c30bd111f2a070.jsonl", "id_parag": "2310.19359-66c30bd111f2a070_3", "final_parag": "\\textbf{Visualizing the predictions}. \nIn the last paragraph, we have analyzed how the patch-level predictions are distributed inside non-cancerous images \\emph{quantitatively}. Indeed, Figure \\ref{fig:negative_bags} represents each patch through a bar. \nHowever, this hampers the \\emph{qualitative} visualization from a medical viewpoint. Here we focus on such qualitative assessment by visualizing the predictions obtained for image 16B0028138, see Figure \\ref{fig:visual_preds}. We have chosen this image because it illustrates best the effect of the coupling term. \nNotice that, thanks to these terms, the proposed VGPMIL-PR-I manages to keep all patches with a probability closer to zero than VGPMIL and VGPMIL-PR. As a consequence, VGPMIL-PR-I is the only method that correctly classifies this image as non-cancerous, recall Figure \\ref{fig:negative_bags}. \nFor the other two methods, there are some patches that trigger the image prediction to be cancerous.", "candidate_comments": ["Interestingly, the correlation introduced by the Ising model can help at this, since they contribute to keep patch-level predictions more homogeneous."]}
{"annotations": [{"id": 86904026, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86971047, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86970915, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19359-66c30bd111f2a070.jsonl", "id_parag": "2310.19359-66c30bd111f2a070_4", "final_parag": "\\textbf{An explicit analysis on the role of $\\lambda$}. \nThe hyperparameter $\\lambda$ is at the core of the novel VGPMIL-PR-I. \nIt was introduced in the probabilistic model to regulate the strength of the coupling term, recall eq.~\\eqref{eq:coupling_term}. \nThis role has been confirmed \\emph{indirectly} in Table \\ref{tab:cms}: when $\\lambda$ gets higher, the predictive performance improves for negative bags and degrades for positive ones. This can be explained because a higher $\\lambda$ homogenizes the patch-level predictions and difficulties the appearance of positive patches. Here we perform a more \\emph{direct} measure to gain insights into the role of $\\lambda$. \nSpecifically, we define the ``variability inside a bag'' as the standard deviation of the probability of cancer for all the patches inside that bag. Therefore, this metric measures the dispersion in the patch-level predictions obtained inside a bag. \nFigure \\ref{fig:lambda_explicit} shows the evolution of this metric for VPGMIL-PR-I as $\\lambda$ increases.", "candidate_comments": ["Interestingly, the correlation introduced by the Ising model can help at this, since they contribute to keep patch-level predictions more homogeneous."]}
{"annotations": [{"id": 86730866, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86731562, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86826520, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19390-a36f1cab88ec0aed.jsonl", "id_parag": "2310.19390-a36f1cab88ec0aed_1", "final_parag": "\\textbf{Rotated MNIST.} For IMGP, we use $\\nu = 2$. \nAs discussed in~\\Cref{appdx:dumbbell}, the optimal number of eigenpairs $L$ varies considerably depending on the convergence of the optimization problem. \nGiven the limited number of iterations per run we opted to fix the number of eigenpairs at $20\\%N$ and $2\\%N$, for SR-MNIST and MR-MNIST, respectively. \nTable \\ref{tab:mnist_results} reports the complete results obtained for the rotated MNIST dataset, including both RMSE and NLL. \nNotably, these emphasize the importance of unlabeled points, as S-IMGP performs worse than both SS-IMGP and EGP on MR-MNIST-1\\%.", "candidate_comments": ["As discussed in~\\Cref{sec:real_examples}, we encountered challenges when seeking appropriate sparse eigen-solvers in the Python ecosystem. First, "]}
{"annotations": [{"id": 86962012, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973149, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86974805, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19390-baba4d42c8bb5065.jsonl", "id_parag": "2310.19390-baba4d42c8bb5065_1", "final_parag": "\\begin{tabular}{cccccccc} \\toprule & \\multicolumn{3}{c}{\\textbf{RMSE}} & & \\multicolumn{3}{c}{\\textbf{NLL}} \\\\ \\cline{2-4} \\cline{6-8} \\\\ $n/N$ & \\textbf{S-IMGP (full)} & \\textbf{SS-IMGP (full)} & \\textbf{EGP} & & \\textbf{S-IMGP (full)} & \\textbf{SS-IMGP (full)} & \\textbf{EGP} \\\\ \\midrule 5\\% & $0.27 \\pm 0.02$ & $0.39 \\pm 0.04$ & $\\mathbf{0.24 \\pm 0.02}$ & & $0.64 \\pm 0.83$ & $\\mathbf{-2.48 \\pm 0.08}$ & $-0.80 \\pm 0.02$ \\\\", "candidate_comments": ["\\begin{tabular}{cccccc}     \\toprule     \\multicolumn{1}{c}{} & \\multicolumn{2}{c}{\\textbf{RMSE}} &     & \\multicolumn{2}{c}{\\textbf{NLL}}                                                                                            \\\\     \\cline{2-3} \\cline{4-6}                                                                                                                                                                     \\\\     $n/N$ & \\textbf{S-IMGP} &  \\textbf{EGP}  &  & \\textbf{S-IMGP} & \\textbf{EGP}  \\\\     \\midrule     5\\%  & $0.250 \\pm 0.020$ & $\\mathbf{0.215 \\pm 0.007}$ &  & $0.473 \\pm 0.062$ & $\\mathbf{-0.447 \\pm 0.012}$ \\\\     10\\% & $\\mathbf{0.147 \\pm 0.010}$ & $0.169 \\pm 0.050$ &  & $0.088 \\pm 0.073$ & $\\mathbf{-0.581 \\pm 0.012}$ \\\\     25\\% & $\\mathbf{0.074 \\pm 0.002}$ & $0.196 \\pm 0.027$ &  & $-0.079 \\pm 0.045$ & $\\mathbf{-0.140 \\pm 0.012}$ \\\\     50\\% & $\\mathbf{0.090 \\pm 0.006}$ & $0.311 \\pm 0.057$ &  & $\\mathbf{-0.146 \\pm 0.073}$ & $5.116 \\pm 0.480$ \\\\     \\bottomrule \\end{tabular}"]}
{"annotations": [{"id": 86983881, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86986419, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86986445, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19390-bc3424280b61817f.jsonl", "id_parag": "2310.19390-bc3424280b61817f_1", "final_parag": "Assuming the manifold is connected, by we have $\\abs[0]{f^{\\c{M}}_l} \\leq C_{\\lambda} \\del{\\lambda^{\\c{M}}_l}^{\\frac{d-1}{4}}$ for $l > 0$ \\cite{donnelly2006}. \nThe case $l = 0$ is special because $\\lambda^{\\c{M}}_0 = 0$. \nSince $f^{\\c{M}}_0$ is a constant function, we have \\[ \\abs[0]{f^{\\c{M}}_l} \\leq C_{\\lambda} \\max(\\del[0]{\\lambda^{\\c{M}}_l}^{\\frac{d-1}{4}}, 1) \\] for all $l \\geq 0$, where $C_{\\lambda}$ here is potentially different from the $C_{\\lambda}$ before. \nAssuming, without loss of generality, $\\eps < 1$, we have \\[ \\abs{ f_l(\\v{x}_i)f_l(\\v{x}_j) - f^{\\c{M}}_l(\\v{x}_i) f^{\\c{M}}_l(\\v{x}_j) } &\\leq \\abs{ f_l(\\v{x}_i) } \\abs{ f_l(\\v{x}_j) - f^{\\c{M}}_l(\\v{x}_j) } \\\\ &~~~~+ \\abs{ f^{\\c{M}}_l(\\v{x}_j) } \\abs{ f_l(\\v{x}_i) - f^{\\c{M}}_l(\\v{x}_i) } \\\\ &\\leq \\eps' \\cdot \\del{ \\abs{ f_l(\\v{x}_i) } + \\abs{ f^{\\c{M}}_l(\\v{x}_j) } } \\\\ &\\leq \\eps' \\cdot \\del{ \\abs{ f_l(\\v{x}_i) - f^{\\c{M}}_l(\\v{x}_i) } + \\abs{ f^{\\c{M}}_l(\\v{x}_i) } + \\abs{ f^{\\c{M}}_l(\\v{x}_j) } } \\\\ &\\leq \\eps' \\cdot \\del[1]{ \\eps' + 2 C_{\\lambda} \\max(\\del[0]{\\lambda^{\\c{M}}_l}^{\\frac{d-1}{4}}, 1) } \\leq \\frac{(1 + 2 C_{\\lambda}) \\eps}{L} . \n\\]", "candidate_comments": ["Hence $\\abs{\\Phi(\\lambda_l) - \\Phi(\\lambda^{\\c{M}}_l)} \\leq C_{\\Phi}\\abs{\\lambda_l - \\lambda^{\\c{M}}_l} \\leq C_{\\Phi} \\eps$."]}
{"annotations": [{"id": 86722816, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86732227, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86730002, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19390-bc3424280b61817f.jsonl", "id_parag": "2310.19390-bc3424280b61817f_2", "final_parag": "Define an auxiliary kernel with manifold eigenvalues and graph eigenfunctions by \\[ \\tilde{k}^L_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}, \\v{x}') = \\frac{\\sigma_f^2}{C_{\\nu, \\kappa}} \\sum_{l=0}^{L} \\del{\\frac{2 \\nu}{\\kappa^2} + \\lambda^{\\c{M}}_l}^{-\\nu-\\dim(\\c{M})/2} f_l(\\v{x}) f_l(\\v{x}'). \n\\]", "candidate_comments": ["Write \\[ \\frac{C_{\\nu, \\kappa}}{\\sigma_f^2} \\abs{ k^L_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) - k^{N, \\alpha, L}_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) } \\leq \\sum_{l=0}^L \\del{ \\abs{} } \\]"]}
{"annotations": [{"id": 86903817, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972363, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86972352, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19390-bc3424280b61817f.jsonl", "id_parag": "2310.19390-bc3424280b61817f_3", "final_parag": "Then \\[ \\frac{C_{\\nu, \\kappa}}{\\sigma_f^2} \\abs{ k^L_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) - \\tilde{k}^{L}_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) } &\\leq \\sum_{l=0}^L \\del{\\frac{2 \\nu}{\\kappa^2} + \\lambda^{\\c{M}}_l}^{-\\nu-\\dim(\\c{M})/2} \\frac{(1 + 2 C_{\\lambda}) \\eps}{L} \\\\ &\\leq \\Phi(0) (1 + 2 C_{\\lambda}) \\eps . \n\\] Also, noting that $\\abs{f_l(\\v{x}_i)} \\leq \\abs{f^{\\c{M}}_l(\\v{x}_i) - f_l(\\v{x}_i)} + \\abs{f^{\\c{M}}_l(\\v{x}_i)} \\leq \\eps' + C_{\\lambda} \\max(\\del[0]{\\lambda^{\\c{M}}_l}^{\\frac{d-1}{4}}, 1)$, write \\[ \\frac{C_{\\nu, \\kappa}}{\\sigma_f^2} \\abs{ \\tilde{k}^L_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) - k^{N, \\alpha, L}_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) } &\\leq \\sum_{l=0}^L \\abs{\\Phi(\\lambda^{\\c{M}}_l)-\\Phi(\\lambda_l)} \\abs{f_l(\\v{x}_i)} \\abs{f_l(\\v{x}_j)} \\\\ &\\leq \\sum_{l=0}^L C_{\\Phi} \\abs{\\lambda^{\\c{M}}_l - \\lambda_l} \\abs{f_l(\\v{x}_i)} \\abs{f_l(\\v{x}_j)} \\\\ &\\leq \\sum_{l=0}^L 2 C_{\\Phi} \\eps' \\del{ (\\eps')^2 + C_{\\lambda}^2 \\max(\\del[0]{\\lambda^{\\c{M}}_l}^{\\frac{d-1}{2}}, 1) } \\\\ &\\leq 2 C_{\\Phi} (1 + C_{\\lambda}^2) \\eps . \n\\] Finally, \\[ \\abs{ k_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) - k^{N, \\alpha, L}_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) } \\leq &\\abs{ k_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) - k^L_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) } \\\\ &+ \\abs{ k^L_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) - \\tilde{k}^{L}_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) } \\\\ &+ \\abs{ \\tilde{k}^L_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) - k^{N, \\alpha, L}_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) } \\\\ &\\leq \\eps + \\frac{\\sigma_f^2}{C_{\\nu, \\kappa}} \\del{ \\Phi(0) (1 + 2 C_{\\lambda}) + 2 C_{\\Phi} (1 + C_{\\lambda}^2) } \\eps . \n\\] This proves the claim. \n\\end{proof}", "candidate_comments": ["Write \\[ \\frac{C_{\\nu, \\kappa}}{\\sigma_f^2} \\abs{ k^L_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) - k^{N, \\alpha, L}_{\\nu, \\kappa, \\sigma_f^2}(\\v{x}_i, \\v{x}_j) } \\leq \\sum_{l=0}^L \\del{ \\abs{} } \\]"]}
{"annotations": [{"id": 86736741, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86829461, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86836484, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19403-e1425d772d0e14e8.jsonl", "id_parag": "2310.19403-e1425d772d0e14e8_1", "final_parag": "discarding yes/no questions and questions for which there is a paragraph answer but not a minimal span answer. \nFor both datasets, we train on original or augmented versions of the training set and report performance on the development set. \nAll data statistics are shown in Table \\ref{tab:data-statistics}.\\\\", "candidate_comments": ["consider the sub-task of returning the minimal span that completely answers the question or NULL if there is no minimal span answer."]}
{"annotations": [{"id": 86865099, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86902305, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86903778, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19405-c5753d6454c08eeb.jsonl", "id_parag": "2310.19405-c5753d6454c08eeb_1", "final_parag": "Object detection is a core component of perception systems, providing the ego vehicle with information about its surroundings to ensure safe route planning. While cameras and Lidar have significantly advanced perception systems, their performance can be limited in adverse weather conditions. In contrast, millimeter-wave technology enables radars to function effectively in such conditions. However, relying solely on radar for building a perception system doesn't fully capture the environment due to the data's sparse nature. To address this, sensor fusion strategies have been introduced. We propose a dual-branch framework to integrate radar and Lidar data for enhanced object detection. The primary branch focuses on extracting radar features, while the auxiliary branch extracts Lidar features. These are then combined using additive attention. Subsequently, the integrated features are processed through a novel Parallel Forked Structure (PFS) to manage scale variations. A region proposal head is then utilized for object detection. We evaluated the effectiveness of our proposed method on the Radiate dataset using COCO metrics. The results show that it surpasses state-of-the-art methods by $1.89\\%$ and $2.61\\%$ in favorable and adverse weather conditions, respectively. This underscores the value of radar-Lidar fusion in achieving precise object detection and localization, especially in challenging weather conditions. \n\\end{abstract}", "candidate_comments": ["Learning world models in terms of perception of the environment is instrumental in developing safe autonomous vehicles.  Object detection, a fundamental building block of perception, provides ego vehicle information about the surroundings to plan a safe route. Although cameras and Lidar have contributed to building a propitious perception system, yet have restricted performance in adverse weather. However, millimetre-wave technology allows the radar to be deployed in adverse weather. The standalone usage of radar to build the perception system is counterproductive in fully representing the environment due to the sparse nature of data. To rectify this, sensor fusion strategies are being adopted, and we proposed a two-branched framework to fuse the radar and Lidar data for object detection. The primary branch extracts the radar's features, whereas the auxiliary branch extracts the Lidar's features and combines them with radar features by employing additive attention. Later, the embedded features are passed to the novel parallel forked structure (PFS) to handle the scale variations. Finally, a region proposal head is adopted for object detection. The proposed method efficacy is evaluated on the Radiate dataset using COCO metrics, outperforming the state-of-the-art methods by a margin of $1.89\\%$ and $2.61\\%$ in good and adverse weather conditions, respectively. This shows that   radar-Lidar fusion contributes to accurate object detection and localization in difficult weather conditions."]}
{"annotations": [{"id": 87255402, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87259675, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 87267880, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19405-c5753d6454c08eeb.jsonl", "id_parag": "2310.19405-c5753d6454c08eeb_2", "final_parag": "Environmental perception is crucial for crafting dependable decision-making algorithms in autonomous vehicles \\cite{wen2022deep}. Modern autonomous vehicles are outfitted with a variety of advanced sensors, including Lidars, cameras, and radars \\cite{azam2020system} \\cite{velasco2020autonomous}. While leveraging these diverse sensor modalities enhances accuracy and robustness in environmental perception, it also poses challenges in creating cohesive perception systems for such vehicles. Moreover, many of these sensors face operational constraints \\cite{jiao2023msmdfusion}. Fusing data from multiple modalities offers a potential solution, providing a richer environmental representation and addressing the individual limitations of each sensor \\cite{chen2017multi} \\cite{kuang2020multi} \\cite{zhang2023perception}. \n\\par In the literature, most existing sensor fusion approaches discuss the fusion of Lidar and camera. Although these two sensors give rich information about the environment, yet there are some limitations in using them. For instance, the point cloud from the Lidar gives dense $3$D measurement at a close range but becomes sparse at long ranges, thus limiting its ability to detect objects at a far length \\cite{zhang2022not}. Similarly, cameras provide a rich spatial representation of the environment but do not provide depth information compared to Lidar data. However, the fusion of Lidar and camera complement each other and give promising results in 3D object detection for autonomous driving, both of these sensors are prone to adverse weather conditions, which can significantly reduce their performance \\cite{chen2017multi} \\cite{kuang2020multi}. \nConversely, because radars operate at millimetre wavelength, they can penetrate or diffract around the tiny particles in adverse weather conditions such as fog, snow, rain and dust, thus can be utilized for the long-range perception of the environment \\cite{qian2021robust} \\cite{li2022exploiting} \\cite{li2022modality}. \n\\par Recently, some research efforts have been made to use radar data as a single modality for object detection. Some common approaches that utilize radar data either represent the radar data in image format or as point cloud data. For instance, \\cite{azam2021channel} have adopted the radar image representation and designed a channel boosting approach in an ensemble feature extractor network for feature representation. Later, they utilized the transformer network to learn the contextual representation between the embedded features and used them for object detection. Similarly, \\cite{li2022exploiting} have transformed the radar data into an ego-centric bird-eye-view radar image to exploit the temporal relations and used them for object recognition. However, majority of works have utilized radar data for object detection only as a point cloud representation. That said, these approaches provide promising improvement in radar-based object detection, yet usage of standalone radar for the environment perception is limited due to its low spatial resolution. To address this, radar data can be fused with Lidar and camera to complement each other \\cite{chu2023mt}. The camera-radar fusion improves the spatial resolution of the perception system but results in the sparse representation of the environment.\\cite {nabati2021centerfusion}. Some works exploit this and fuse the radar and Lidar data for object detection \\cite{qian2021robust}. However, sensor fusion between radar and Lidar provides a better representation of the environment, but how to fuse the data between these two modalities is an open research problem. \n\\par", "candidate_comments": ["The perception of the environment plays a critical role in designing safe and reliable decision-making algorithms for the autonomous vehicle \\cite{wen2022deep}. Nowadays, autonomous vehicles are equipped with multiple state-of-the-art sensor modalities, for instance, Lidars, cameras and radars \\cite{azam2020system} \\cite{velasco2020autonomous}. Utilizing these multiple sensor modalities provides accuracy and robustness to perceive the environment yet introduces design challenges to developing unified perception systems for autonomous vehicles. In addition, most of these sensors have limitations in operability \\cite{jiao2023msmdfusion}. To mitigate these issues, fusing the multi-modal data is one possible solution in this direction that not only gives us an enriching representation of the environment but also overcomes the individual sensor's limitations \\cite{chen2017multi} \\cite{kuang2020multi} \\cite{zhang2023perception}."]}
{"annotations": [{"id": 86737426, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86901321, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86901282, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19405-c5753d6454c08eeb.jsonl", "id_parag": "2310.19405-c5753d6454c08eeb_3", "final_parag": "Neural networks allow for diverse sensor data fusion strategies due to their hierarchical feature representation \\cite{feng2020deep}. Commonly, these strategies include early, mid, and late fusion. Early fusion combines data before feature extraction but can face misalignment issues. Late fusion works at the decision level, but may not capture the full potential of joint sensor data representations. Mid-fusion strikes a balance between the two, effectively learning joint representations of sensor modalities \\cite{feng2020deep}. \nSome works have fused the radar data with Lidar and camera to complement the sensor's individual limitations. \\cite{nabati2021centerfusion} have fused the radar and camera by first finding the centre points in image data using the centre point detection network. Then, it solves the data association problem between radar detections and the object's centre using a frustum-based method. Later, the associated radar features are used for radar object detection. Another work, RODNet \\cite{wang2021rodnet} is a cross-supervised camera-radar fusion framework for radar object detection. It uses the teacher-student network, where the teacher network fuses both the radar images and RGB images to obtain the object classes and location in the radar images. Finally, the teacher network supervises the student network that only uses radar images for object detection. MT-DETR \\cite{chu2023mt} has proposed a multi-modal fusion network by introducing the residual and confidence fusion modules to fuse the radar, camera and Lidar. In addition, a residual enhancement module is designed to enhance each unimodal branch's feature extraction. Similarly, \\cite{nobis2021radar} have also fused the Lidar, camera and radar for object detection. MVDNet \\cite{qian2021robust} has fused the Lidar and radar by first generating the proposal from two sensor data and then adopting the region-wise features to improve the object detections. Similarly, ST-MVDNet \\cite{li2022modality} is built on top of MVDNet \\cite{qian2021robust} by introducing the teacher-student network for radar-based object detection. In their framework, the teacher model generates the predictions to train the student network using a consistency constraint while the student network passes the parameters it learns to the teacher network via an exponential moving average. \\par MVDNet \\cite{qian2021robust} and ST-MVDNet \\cite{li2022modality} are more similar to our work employing radar and Lidar fusion for object detection. However, both of these methods only consider foggy weather conditions. In addition, MVDNet \\cite{qian2021robust} adopts a late fusion approach that limits learning the joint distribution of sensor modalities in contrast to our mid-fusion approach. However, ST-MVDNet \\cite{li2022modality} uses the knowledge distillation approach, which has a problem of large discrepancy of predictive distribution between teacher and student networks. To address these issues, we designed a novel fusion network that uses additive attention for sensor fusion and depth-wise convolution to extract useful features from the sparse data of radar and Lidar, respectively. In addition, we introduce a parallel fork structure to handle the scale variations that solve the predictive discrepancy between the learned features.", "candidate_comments": ["The neural network's hierarchical feature representation permits the design of multiple strategies for fusing the sensor data \\cite{feng2020deep}. The most common approaches that have been adopted fall in early, late or mid-fusion schemes. In the early fusion, the sensor data are fused at the data level before extracting the features. This scheme helps the feature extractor network learn the sensor modalities' joint representation. However, the early fusion strategies are sensitive to spatial and temporal misalignment of the data \\cite{feng2020deep}. On the other hand, late fusion approaches fuse the sensor's data at a decision level and, thus, therefore provide more room for introducing new sensor modalities to the feature extraction network. The common problem in the late fusion scheme is that the full potential of sensor modalities is not exploited because the joint representation between the sensor's data is not appropriately learnt \\cite{feng2020deep}. However, the mid-fusion compromises early and late fusion, enabling the network to learn the joint representation between the sensor modalities \\cite{feng2020deep}. "]}
{"annotations": [{"id": 86952083, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86959849, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}, {"id": 86972749, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19415-8a8f492ff7bdf707.jsonl", "id_parag": "2310.19415-8a8f492ff7bdf707_1", "final_parag": "Recently, techniques based on Score Distillation Sampling (SDS)~\\citep{poole2022dreamfusion,lin2023magic3d,chen2023fantasia3d,wang2023prolificdreamer}, also known as Score Jacobian Chaining (SJC)~\\citep{wang2023score}, have emerged as a major research direction for text-to-3D generation, as they can produce high-quality and intricate 3D results from diverse text prompts without requiring 3D data for training. The core principle behind SDS is to optimize 3D representations by encouraging their rendered images to move towards high probability density regions conditioned on the text, where the supervision is provided by a pre-trained 2D diffusion model \\citep{ho2020denoising,sohl2015deep,rombach2022high,saharia2022photorealistic,balaji2022ediffi}. \nDreamFusion \\citep{poole2022dreamfusion} advocates the use of SDS for the optimization of Neural Radiance Fields (NeRF). \\citep{barron2022mip,mildenhall2021nerf}. Subsequent research improve the visual quality by introducing coarse-to-fine optimization strategies \\citep{lin2023magic3d,chen2023fantasia3d,wang2023prolificdreamer}, efficient 3D representations \\citep{lin2023magic3d,chen2023fantasia3d},", "candidate_comments": ["Recently, methods based on score distillation sampling (SDS) \\xjqi{add citations in this line of research} have become one major stream of research since they do not need 3D data for training but can easily produce high-quality and intricate 3D results from diverse text prompts.  The core idea of SDS is to optimize a NeRF by encouraging images rendered from it to move toward high conditional data density regions during training, wherein the conditional data density estimation is from a 2D diffusion model \\xjqi{check}. Many follow-up works improve the visual quality by introducing coarse-to-fine strategies \\xjqi{add citation}, geometry (e.g. normal) renderings \\xjqi{add citation}, or new perspective of parameter optimization, treating 3D parameters of NeRF as random variables \\xjqi{add citation}. Albeit advancing the area, they all rely on SDS for optimization.   is to employ the generative prior learned by a 2D diffusion model to supervise optimizing a 3D model represented as a NeRF through drawing images rendered from the NeRF toward high-density ", "Recently, methods based on score distillation sampling have become one major stream since they do not need 3D data for training. The key idea is utilizing a pretrained text-to-image diffusion model as a supervision for the underlying 3D representation. For instance, Dreamfusion first introduces score-distillation sampling to optimize a neural radiance field (NeRF). Magic3D further leverages a coarse-to-fine strategy to get superior mesh quality and higher-resolution texture results. Fantasia3D proposes to use SDS to supervise the surface normal and get a better geometry result. Most recently, ProlificDreamer integrated Variational Score Distillation that treats the 3D parameter as a random variable instead of a constant, which greatly improves the over-saturation, over-smoothing, and low-diversity problems. While using different strategies or representations in these methods, the central part of these methods is the effectiveness of generative prior learned by the pretrained diffusion model, which provides an update direction for the rendered images to move to a higher-density region.  "]}
{"annotations": [{"id": 86829837, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}, {"id": 86856194, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}, {"id": 86898470, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19415-8a8f492ff7bdf707.jsonl", "id_parag": "2310.19415-8a8f492ff7bdf707_2", "final_parag": "In this paper, we show that the classifier component of Score Distillation Sampling (SDS) is not just auxiliary but essential for text-to-3D generation, and using only this component is sufficient. We call this paradigm \\textit{Classifier Score Distillation (CSD)}. \\textit{This insight fundamentally shifts our understanding of the mechanisms underlying the success of text-to-3D generation based on score distillation. Specifically, the efficacy stems from distilling knowledge from an implicit classifier $p_\\phi\\left(y|\\mathbf{x}_t\\right)$ rather than from reliance on the generative prior $p_\\phi\\left(\\mathbf{x}_t|y\\right)$}. This finding further enables us to uncover fresh insights into existing design choices: 1) We demonstrate that utilizing negative prompts can be viewed as a form of joint optimization with dual classifier scores. This realization enables us to formulate an annealed negative classifier score optimization strategy, which enhances generation quality while maintaining the result faithfulness according to prompts. 2) Building on our understanding, we illustrate the utilization of classifier scores for efficient text-driven 3D editing. 3) The Variational Score Distillation technique \\citep{wang2023prolificdreamer} can be viewed as an adaptive form of negative classifier score optimization, where the negative direction is supplied by a diffusion model trained concurrently.", "candidate_comments": ["We find that removing the generative component is not harmful to optimization, and the discriminative component is sufficient for text-to-3D generation, which we call classifier score distillation. This new perspective of understanding SDS in practice further sheds light on new insights into existing design choices and enables us to develop new strategies: 1) We demonstrate that the use of negative prompts can be viewed as a form of joint optimization with dual classifier scores, enabling us to develop an annealed negative classifier score optimization strategy. This strategy can improve the quality while keeping the faithfulness of the original prompt; 2)  The new Variational Score Distillation method can be viewed as an adaptive form of negative classifier score optimization, where the negative direction is provided by the parallel-trained diffusion model \\xjqi{LoRA adapted image}, and thus it does not need a very large guidance scale; 3) We develop new classifier scores that enable to us to edit 3D scenes.", "In this paper, we show that solely relying on the discriminative part is not only feasible but effective for text-to-3d generation tasks, which we call classifier score distillation. This new finding enables us to present a new perspective for understanding some existing techniques and developing new strategies. Specifically, we demonstrate that the use of negative prompts can be viewed as a form of joint optimization with dual classifier scores, enabling the development of an annealed negative classifier score optimization strategy, which can improve the quality while keeping the faithfulness of the original prompt. Furthermore, Variational Score Distillation appears as an adaptive form of negative classifier score optimization, where the negative direction is provided by the parallel-trained diffusion model, and thus it does not need a very large guidance scale. Lastly, we show how to use classifier scores to edit 3D scenes."]}
{"annotations": [{"id": 86962857, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86965271, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86970699, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19424-0134def8df2a6427.jsonl", "id_parag": "2310.19424-0134def8df2a6427_1", "final_parag": "Our work enables agents to reach any previously visited states by learning goal-conditioned policies that cover the entire goal space. In contrast, exploration bonuses help agents visit novel states, but they cannot reuse learned policies to solve user-specified goals as those states are quickly forgotten.", "candidate_comments": ["While exploration bonus helps agents to visit novel states, they can not reuse learned policies to solve user-specified goals as previously visited novel states are forgotten quickly. In contrast, our work enables to command the agent to reach any states that it visited by learning the goal-conditioned policy where the goal provably covers the entire goal space."]}
{"annotations": [{"id": 86972388, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86982084, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86982038, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19424-54d4c018414d9852.jsonl", "id_parag": "2310.19424-54d4c018414d9852_1", "final_parag": "As described in the prior work \\cite{warde-farley2019unsupervised, choi2021variational}, it has been shown that Equation \\ref{eq:policy_objective} which is also called an intrinsic reward \\cite{gregor2016variational}, recovers the objective of GCRL in Equation \\ref{eq:gcrl_objective} with dense rewards. By choosing a Gaussian distribution with mean $s$ and fixed variance $\\sigma^2 I$ for $q_\\lambda(g|s)$ where $I$ is the identity matrix, this objective becomes a negative $l_2$ distance between $s$ and $g$. Similarly, one can show that the intrinsic reward represented in Equation \\ref{eq:policy_objective} becomes a sparse reward where an agent gets $0$ reward if $l_2$ distance between $s$ and $g$ is within some threshold $\\delta_g$ and gets $-1$ otherwise. Other MI-based methods can also be considered a GCRL by modeling $q_\\lambda(g|s)$ to follow $\\mathcal{N}(\\mu(s), \\sigma^2 I)$ where $\\mu(s)$ is a function approximator usually following an encoder structure.", "candidate_comments": ["Similarly, by formulating $\\frac{1}{Z} \\exp(1-2\\delta_g\\mathcal{U}_{[g-\\delta_g, g+\\delta_g]})$ for the discriminator, one can show that the intrinsic reward becomes a sparse reward where an agent gets $0$ reward if $l_2$ distance between $s$ and $g$ is within some threshold $\\delta_g$ and gets $-1$ otherwise. Other MI-based methods can also be considered as a goal-conditioned RL by modeling $q_\\lambda(g|s)$ to follow $\\mathcal{N}(\\mu(s), \\sigma^2 I)$ where $\\mu(s)$ is a function approximator usually with an encoder structure. \na specific distribution as shown in Table \\ref{tab:vcrl}."]}
{"annotations": [{"id": 86971168, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86973035, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974979, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19424-6e9f2a86eec1207b.jsonl", "id_parag": "2310.19424-6e9f2a86eec1207b_1", "final_parag": "We also demonstrate that VUVC enables a real-world robot to learn to navigate in a long-range environment without any explicit rewards, and that incorporating skills with a global planner further improves the performance.", "candidate_comments": ["VUVC also enables a real-world robot to learn to navigate in the long-range environment without any explicit rewards. On the real-world task, we illustrate that goal-conditioned skills incorporated with a global planner can even further improve the performance."]}
{"annotations": [{"id": 86725416, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86729625, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86730214, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19424-6e9f2a86eec1207b.jsonl", "id_parag": "2310.19424-6e9f2a86eec1207b_2", "final_parag": "We provide the unifying framework VCRL which recasts MI-based RL as curriculum learning in goal-conditioned RL. Under VCRL framework, we propose a novel approach VUVC for unsupervised discovery of skills which utilizes a value uncertainty for an increment in the entropy of the visited state distribution. Under regularity conditions, we prove that VUVC improves the expected entropy more than the uniform curriculum method. Our experimental results demonstrate that VUVC consistently outperforms a variety of prior methods both on configuration-based and vision-based continuous robot manipulation tasks.", "candidate_comments": ["Contributions - Providing VCRL framework which encapsulates existing  - Under this framework, we propose a novel approach for unsupervised discovery of skills which ... \n- Prove that under regularity conditions, ... \n- Demonstrate that VUVC consistently outperforms a variety of prior methods in ~ env and ~ env, even in real-robot. \n- In real-world robot experiments, we demonstrate how our method can be utilized for long-range navigation tasks / transferred to unseen environment well."]}
{"annotations": [{"id": 86961878, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86977018, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86978357, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent"]}], "agreement": 55.55555555555555, "source_file": "2310.19424-7c8f6a4ca5bb0ed7.jsonl", "id_parag": "2310.19424-7c8f6a4ca5bb0ed7_1", "final_parag": "Especially in the complicated \\textit{PointMazeSquareLarge} environment, VUVC requires much less interaction for task completion. This result suggests the importance of an elaborate curriculum goal distribution in comparison to GoalGAN or Skew-Fit and emphasizes the importance of a gradually increasing state covering goal distribution when compared to EDL and EDL-Oracle.", "candidate_comments": ["The result in \\textit{PointMazeSquareLarge} which is the most complicated environment among point mazes suggests the importance of a gradually increasing state covering goal distribution. In this environment, VUVC requires much less interactions than EDL and EDL-Oracle for task completion. Unlike EDL, VUVC generates goals from a non-stationary goal distribution considering the ability of the agent, which encourages the agent to achieve meaningful learning signal. \nHowever, EDL gives curriculum goals uniformly from the entire state space sampled from a stationary distribution, some of which are hard to reach in the early phase of training.", "In addition, EDL-Oracle has large standard deviation over random seeds in both FetchPickAndPlace task and FetchSlide task (see red area), which also suggests the importance of a gradually increasing goal distribution."]}
{"annotations": [{"id": 86952128, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86961201, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86959118, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19424-7c8f6a4ca5bb0ed7.jsonl", "id_parag": "2310.19424-7c8f6a4ca5bb0ed7_2", "final_parag": "To see the effects of the value uncertainty in the curriculum, in Figure \\ref{fig:prop}, we investigate (1) how the value uncertainty $U(g)$ and log density of visited states $p_t^{\\mathrm{visited}}$ are correlated, and (2) how well the value uncertainty filters out uninformative states. In general, we observe that $U(g)$ and $p_t^{\\mathrm{visited}}$ show negative correlation, indicating that VUVC covers the state space faster than the uniform curriculum for the optimal goal-conditioned policy as the regularity condition of Proposition \\ref{prop_2} holds empirically. \nIn addition to this, we observe a case which satisfies the regularity condition of Proposition \\ref{prop_3} from the landscape visualization, implying that our method is more effective than the uniform curriculum.", "candidate_comments": ["In addition to this, the landscape visualization of the value uncertainty and success rate demonstrates that our curriculum satisfies the regularity condition of Proposition \\ref{prop_3}."]}
{"annotations": [{"id": 86720830, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86724845, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86729420, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19424-a580c5776e854945.jsonl", "id_parag": "2310.19424-a580c5776e854945_1", "final_parag": "\\begin{proof} We substitute the $\\mathcal{G}=\\mathcal{G}_{\\mathrm{exploit}} \\cup \\mathcal{G}_{\\mathrm{uninfo}} \\cup \\mathcal{G}_{\\mathrm{info}}$ into the entropy increment over uniform curriculum and expand the expression:", "candidate_comments": ["\\begin{proof} We substitute the $\\mathcal{G}=\\mathcal{G}_{\\mathrm{exploit}} \\cup \\mathcal{G}_{\\mathrm{explore}}$ into the entropy increment over uniform curriculum and expand the expression:"]}
{"annotations": [{"id": 86830054, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86856375, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86898755, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19424-a580c5776e854945.jsonl", "id_parag": "2310.19424-a580c5776e854945_2", "final_parag": "\\item{\\textbf{Reinforcement learning with Imagined Goals (RIG)} \\cite{nair2018visual}}: RIG trains a goal-conditioned policy in an unsupervised manner by estimating the visited state distribution and automatically setting curriculum goals sampled from this distribution.", "candidate_comments": ["Instead of making use of a stationary state distribution induced by the current goal-conditioned policy for approximating the state distribution like RIG,"]}
{"annotations": [{"id": 86863827, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86959888, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86960637, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19424-a580c5776e854945.jsonl", "id_parag": "2310.19424-a580c5776e854945_3", "final_parag": "\\item{\\textbf{Explore, Discover and Learn (EDL)} \\cite{campos2020explore}}: EDL overcomes the limitation of existing variational empowerment methods which provide a poor coverage of the state space. \nUnlike RIG which makes use of the current goal-conditioned policy to approximate the state distribution,", "candidate_comments": ["Instead of making use of a stationary state distribution induced by the current goal-conditioned policy for approximating the state distribution like RIG,"]}
{"annotations": [{"id": 86982174, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86983071, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86983707, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19424-aa54d29279188aaf.jsonl", "id_parag": "2310.19424-aa54d29279188aaf_1", "final_parag": "Empowerment or MI-based RL \\cite{klyubin2005empowerment, salge2014empowerment} has gained traction in recent years as a means of unsupervised skill discovery due to its intuitive interpretation and empirical successes \\cite{eysenbach2018diversity, sharma2019dynamics, jabri2019unsupervised}. However, the common empowerment approach has been to either fix or parameterize the distribution of skills \\cite{nair2018visual, pong2020skew, campos2020explore}. The efficiency of learning skills with respect to the number of required training samples is rather limited when the agent learns complex skills from a fixed skill distribution without an organized order. \nThe notion of \\textit{curriculum} studies the effectiveness of the order of training skills. By selecting the order of appropriate skills, a learning agent may achieve a variety of complex skills \\cite{florensa2018automatic, fang2019curriculum}. However, it is both necessary to define a set of tasks that can be used to generate curriculum \\cite{klink2020self, zhang2020automatic} and specify a form of reward functions \\cite{racaniere2019automated, ren2019exploration, narvekar2019learning}.", "candidate_comments": ["Although the notion of \\textit{curriculum} studies the effectiveness of the order of training skills \\cite{florensa2018automatic, fang2019curriculum}, it usually requires a specific form of reward function \\cite{racaniere2019automated, ren2019exploration, narvekar2019learning}."]}
{"annotations": [{"id": 86983388, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986132, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86986169, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19424-aa54d29279188aaf.jsonl", "id_parag": "2310.19424-aa54d29279188aaf_2", "final_parag": "The main contributions of this paper can be summarized as follows: (1) We provide the unifying framework VCRL encapsulating most of the prior MI-based approaches. (2) We propose VUVC, a value uncertainty based approach to information-theoretic skill discovery, aimed at automatically generating curricula for training skills and which is supported by theoretical justification. (3) We show the effectiveness of our approach on complex navigation, robotic manipulation in both configuration and image state space, and real-world robotic navigation tasks and illustrate that the skills discovered by our method can be further improved by incorporating them with a global planner.", "candidate_comments": ["The main contributions of this paper can be summarized as follows: (1) We provide the unifying framework VCRL encapsulating most of the prior MI-based approaches. (2) We propose VUVC, a value uncertainty based approach to information-theoretic skill discovery for automatic curriculum for training skills and prove that under regularity conditions, VUVC increases the expected entropy more compared to uniform curriculum. (3) We show the effectiveness of our approach on complex navigation, robotic manipulation in state space based both configuration and image, and real-world robotic navigation tasks and illustrate that the skills discovered by our method can be further improved by incorporating them with a global planner."]}
{"annotations": [{"id": 86865152, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}, {"id": 86901502, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86901696, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19424-ece6fd487f577716.jsonl", "id_parag": "2310.19424-ece6fd487f577716_1", "final_parag": "We quantify the uncertainty of value functions in predictions of the ensemble members from the initial state by computing the variance over the ensemble of the value functions:", "candidate_comments": ["To quantify the uncertainty of value functions, we use the disagreement in predictions of an ensemble members from the initial state, which is computed as the variance over an ensemble of value functions:", "We quantify the uncertainty of value functions in predictions of an ensemble members from the initial state, which is computed as the variance over an ensemble of value functions:"]}
{"annotations": [{"id": 86982128, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86984145, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86984419, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19424-ece6fd487f577716.jsonl", "id_parag": "2310.19424-ece6fd487f577716_2", "final_parag": "\\begin{restatable}{proposition}{Firstprop} If $V_{\\psi}(s_0, g)$ follows a log-concave distribution, then we have \\end{restatable}", "candidate_comments": ["\\begin{restatable}{proposition}{Firstprop} If $V_{\\psi}(s_0, z)$ follows a log-concave distribution, then selecting a skill/goal maximizing the disagreement in predictions of an ensemble of value functions is equivalent to maximizing the lower bound approximation of expected information gain: \\end{restatable}"]}
{"annotations": [{"id": 86904020, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972344, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86972372, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19424-ece6fd487f577716.jsonl", "id_parag": "2310.19424-ece6fd487f577716_3", "final_parag": "We rewrite the mutual information as the difference between conditional entropy and marginal entropy. We then use the result in \\cite{marsiglietti2018lower} on a lower bound on the entropy of a log-concave random variable, expressed in terms of the $p$-th absolute moment to obtain the conclusion. The complete proof appears in Appendix \\ref{A_Proofs}. \n\\end{sproof}", "candidate_comments": ["\\textcolor{kyowoon_color}{We base on the result in \\cite{marsiglietti2018lower} on the lower bound on the entropy of a log-concave random variable. And ....}  \\textcolor{kyowoon_color}{We rewrite the mutual information as conditional entropy subtracted from marginal entropy. We then use the result in \\cite{marsiglietti2018lower} on a lower bound on the entropy of a log-concave random variable derived in terms of $p$-th absolute moment to obtain the conclusion. The complete proof appears in Appendix \\ref{A_Proofs}.} "]}
{"annotations": [{"id": 86829504, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86861711, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86898825, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19424-ece6fd487f577716.jsonl", "id_parag": "2310.19424-ece6fd487f577716_4", "final_parag": "We term a VCRL method with a goal generative model following Equation \\ref{eq:vuvc_probs} as VUVC.", "candidate_comments": ["We term a VCRL method whose goal generative model follows the distribution in Equation \\ref{eq:vuvc_probs} as Value Uncertainty Variational Curriculum (VUVC)."]}
{"annotations": [{"id": 86736230, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86901580, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86903312, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19424-ece6fd487f577716.jsonl", "id_parag": "2310.19424-ece6fd487f577716_5", "final_parag": "Given the empirical distribution of the visited state [EQUATION] where $\\mathbb{I}(\\cdot)$ is an indicator function, uniform curriculum goal distribution $p_t^{\\mathcal{U}}$ and value uncertainty-based curriculum goal distribution $p_t^{\\mathrm{VU}}$ are defined as follows:", "candidate_comments": ["Given the empirical distribution of the visited state $p_t^{\\mathrm{visited}}$, our curriculum goal distribution $p_t^{\\mathrm{VU}}$ and uniform curriculum goal distribution $p_t^{\\mathcal{U}}$ as "]}
{"annotations": [{"id": 86861753, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86899396, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86899501, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19424-ece6fd487f577716.jsonl", "id_parag": "2310.19424-ece6fd487f577716_6", "final_parag": "We begin by deriving a next step empirical distribution of the visited state given a curriculum goal $g$ and a stationary state distribution induced by the policy $\\rho^{\\pi_{\\theta}}(s|g)$, which can be written as $p_{t+1}^{\\mathrm{visited}}(s)=\\frac{p_{t}^{\\mathrm{visited}}(s) + \\epsilon\\rho^{\\pi_{\\theta}}(s|g)}{1+\\epsilon}$. Plugging this back into Definition \\ref{def:def_1}, we analyze asymptotic behavior of the expected entropy increment and obtain the conclusion with the assumption $\\rho^{\\pi_{\\theta}}(s|g)=\\mathbb{I}(s=g)$. The complete proof is provided in Appendix \\ref{A_Proofs}. \n\\end{sproof}", "candidate_comments": ["\\textcolor{kyowoon_color}{We base on the result in \\cite{marsiglietti2018lower} on the lower bound on the entropy of a log-concave random variable. And ....}  \\textcolor{kyowoon_color}{We begin by deriving a next step empirical distribution of the visited state given a curriculum goal $g$ and a stationary state distribution induced by the policy $\\rho^{\\pi_{\\theta}}(s|g)$, which can be written as $p_{t+1}^{\\mathrm{visited}}(s)=\\frac{p_{t}^{\\mathrm{visited}}(s) + \\epsilon\\rho^{\\pi_{\\theta}}(s|g)}{1+\\epsilon}$. Plugging this back into Definition \\ref{def:def_1}, we analyze asymptotic behavior of the expected entropy increment and obtain the conclusion with the assumption $\\rho^{\\pi_{\\theta}}(s|g)=\\mathbb{I}(s=g)$. See Appendix \\ref{A_Proofs} for the complete proof.} "]}
{"annotations": [{"id": 86982150, "completed_by": {"id": 140156}, "result": ["Différent", "Différent"]}, {"id": 86982201, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}, {"id": 86903804, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19424-ece6fd487f577716.jsonl", "id_parag": "2310.19424-ece6fd487f577716_7", "final_parag": "Define the set $\\mathcal{G}=\\mathcal{G}_{\\mathrm{exploit}} \\cup \\mathcal{G}_{\\mathrm{uninfo}} \\cup \\mathcal{G}_{\\mathrm{info}}$ and positive constant $\\Delta_1, \\Delta_2$ where for all $g \\in \\mathcal{G}_{\\mathrm{uninfo}}$, and for all $g \\in \\mathcal{G}_{\\mathrm{info}}$, Given $\\epsilon=\\frac{1}{t}$, if", "candidate_comments": ["% Define the set $\\mathcal{G}=\\mathcal{G}_{\\mathrm{uninfo_1}} \\cup \\mathcal{G}_{\\mathrm{uninfo_2}} \\cup \\mathcal{G}_{\\mathrm{info}}$ and positive constant $\\Delta$ where % and for all $g \\in \\mathcal{G}_{\\mathrm{explore}}$,", "Define the set $\\mathcal{G}=\\mathcal{G}_{\\mathrm{exploit}} \\cup \\mathcal{G}_{\\mathrm{explore}}$ and positive constant $\\Delta$ where and for all $g \\in \\mathcal{G}_{\\mathrm{explore}}$, Given $\\epsilon=\\frac{1}{t}$, if "]}
{"annotations": [{"id": 86864070, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86960868, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86959541, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19424-ece6fd487f577716.jsonl", "id_parag": "2310.19424-ece6fd487f577716_8", "final_parag": "The proof proceeds in a similar manner as Proposition \\ref{prop_2} except for an assumption $\\mathcal{G}=\\mathcal{G}_{\\mathrm{exploit}} \\cup \\mathcal{G}_{\\mathrm{uninfo}} \\cup \\mathcal{G}_{\\mathrm{info}}$. The complete proof is in Appendix \\ref{A_Proofs}. \n\\end{sproof}", "candidate_comments": ["\\textcolor{kyowoon_color}{We base on the result in \\cite{marsiglietti2018lower} on the lower bound on the entropy of a log-concave random variable. And ....}  \\textcolor{kyowoon_color}{The proof proceeds in a similar manner as Proposition \\ref{prop_2} except for an assumption on the set $\\mathcal{G}=\\mathcal{G}_{\\mathrm{exploit}} \\cup \\mathcal{G}_{\\mathrm{uninfo}} \\cup \\mathcal{G}_{\\mathrm{info}}$. See Appendix \\ref{A_Proofs} for the complete proof.} "]}
{"annotations": [{"id": 86979710, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}, {"id": 86986014, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86986242, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19424-ece6fd487f577716.jsonl", "id_parag": "2310.19424-ece6fd487f577716_9", "final_parag": "Proposition \\ref{prop_3} extends Proposition \\ref{prop_2} to the case where the goal-conditioned policy is sub-optimal and fails to achieve some of the goals. It implies that we need a curriculum method which can filter out uninformative states when the policy can not consistently achieve certain states, in order to achieve a rapid increment of entropy. Empirical observations indicate that VUVC achieves this effect (further details provided in Section \\ref{sec:experiments}).", "candidate_comments": ["(see more details in Section \\nameref{sec:impact_of_uncertainty}). \nThis states that we need a curriculum method which can also filter out states that the policy can consistently achieve, which we empirically observe that VUVC does. \nThis states that we need a curriculum method which can also filter out states that the policy can consistently achieve, which we empirically observe that VUVC does. \nIt states that we need curriculum method which can also filter out where policy can consistently achieve, which we empirically observe that VUVC does.", "It implies that we need a curriculum method which can filter out uninformative states when the policy consistently fails to obtain some states of rapid increment of entropy, as demonstrated in the empirical results of VUVC"]}
{"annotations": [{"id": 86960248, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86972969, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973460, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-05b6758fa4abdb48.jsonl", "id_parag": "2310.19427-05b6758fa4abdb48_1", "final_parag": "To assess the quality of plans generated by diffusion probabilistic models, we propose a novel metric named \\textit{restoration gap}. It aims to automatically detect infeasible plans that violate system constraints. We hypothesize that for feasible plans, even if a certain amount of noise perturbs them, they can be closely restored to their initial plans by diffusion models. It is attributed to the property of temporal compositionality in diffusion planners \\cite{janner2022planning} that encourages them to compose feasible trajectories by stitching together any feasible plan subsequences. However, for infeasible plans that obviously fall out of the training distribution as they violate physical constraints as shown in Figure \\ref{fig:attribution_maps}, restoring them to a state near their original conditions is challenging. Based on this intuition, we define the restoration gap of the generated plan $\\boldsymbol{\\tau}$ as follows: where ${\\hat{t}} \\in (0, 1]$ indicates the magnitude of applied perturbation. The restoration gap measures the expected $L_2$ distance between the generated plan and the plan restored from noise corruption, which is estimated by the Monte Carlo approximation.", "candidate_comments": ["In order to assess the quality of plans generated by diffusion probabilistic models, we propose a novel metric named \\textit{restoration gap}. Our hypothesis is that for feasible plans, even if they are perturbed by a certain amount of noise, they can be closely restored to their initial plans by diffusion models. In contrast, for infeasible plans that are highly likely to fail during execution, it is challenging to restore them to a state near their original conditions. This is attributed to the property of temporal compositionality in diffusion planners \\cite{janner2022planning} which allows them to compose out-of-distribution trajectories by stitching together any feasible plan subsequences. While infeasible plans, especially those that cannot be executed due to system constraints, become inherently harder to restore. Based on this intuition, we define the restoration gap of the generated plan $\\boldsymbol{\\tau}$ as follows:"]}
{"annotations": [{"id": 86899159, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86901428, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86907306, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19427-05b6758fa4abdb48.jsonl", "id_parag": "2310.19427-05b6758fa4abdb48_2", "final_parag": "The histogram of the restoration gap for normal and artifact plans demonstrates that infeasible artifact plans have larger restoration gap values compared to normal plans.", "candidate_comments": ["Considering this, we can standardize the statistic using the restoration gap and identify an artifact plan whenever the restoration gap exceeds the threshold $b > 0$:"]}
{"annotations": [{"id": 86961017, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86973483, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973939, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-05b6758fa4abdb48.jsonl", "id_parag": "2310.19427-05b6758fa4abdb48_3", "final_parag": "Therefore, the detection of infeasible artifact plans can be automated by incorporating a statistical test that utilizes the restoration gap and thresholding with a threshold value of $b > 0$:", "candidate_comments": ["Considering this, we can standardize the statistic using the restoration gap and identify an artifact plan whenever the restoration gap exceeds the threshold $b > 0$:"]}
{"annotations": [{"id": 86910587, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86959313, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86959977, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19427-1f3390c34e419d1d.jsonl", "id_parag": "2310.19427-1f3390c34e419d1d_1", "final_parag": "it entirely relies on the ability of a generative model and assumes a perfect data score estimation.", "candidate_comments": ["it solely employs conditional guidance with gradients of return in the denoising process and assumes a perfect data score estimation."]}
{"annotations": [{"id": 86960528, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86960843, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86962056, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19427-1f3390c34e419d1d.jsonl", "id_parag": "2310.19427-1f3390c34e419d1d_2", "final_parag": "Although guiding the diffusion planner to minimize the restoration gap effectively refines low-quality plans (more details in Section \\ref{sec:experiments}), this refining guidance could push the plan in an undesirable direction due to the estimation error of the gap predictor during the denoising process. As a result of this estimation error, guiding plans with the sub-optimal gap predictor may result in \\textit{model exploitation} \\cite{kurutach2018model, janner2019trust, rajeswaran2020game}, yielding sub-optimal results.", "candidate_comments": ["Through our experiments below, we find that RGG is effective in refining low-quality plans by directing the reduction of the estimated restoration gap. However, we also identify that solely depending on the restoration gap can be limiting when working with offline datasets that lack sufficient diversity. Specifically, if the dataset does not cover the entire state-action space, the learned gap predictor might not be globally accurate. As a result, guiding with a learned model without considering potential inaccuracies may result in \"model exploitation\" \\cite{kurutach2018model, janner2019trust, rajeswaran2020game}, yielding sub-optimal results. "]}
{"annotations": [{"id": 86868642, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86901025, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86901527, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-2cf1ada74b2f65c2.jsonl", "id_parag": "2310.19427-2cf1ada74b2f65c2_1", "final_parag": "\\caption{Performance of plans chosen from the top-K $\\%$ considering various metrics, in the Maze2D-Large single-task. }", "candidate_comments": ["\\caption{Performance of Maze2D-Large single-task, considering up to the top-K $\\%$ plans generated by Diffuser with respect to negative realism score, rarity score, and restoration gap.}"]}
{"annotations": [{"id": 86985323, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985368, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86985378, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-2cf1ada74b2f65c2.jsonl", "id_parag": "2310.19427-2cf1ada74b2f65c2_2", "final_parag": "We evaluate how effectively the restoration gap can identify infeasible plans by comparing our metric with a realism score \\cite{kynkaanniemi2019improved} and rarity score \\cite{han2022rarity}. Both prior metrics are designed to assess the quality of generated samples by examining the discrepancy between the generated sample and the real data manifold in the feature space. Figure \\ref{fig:top_k} illustrates the performance of plans which are chosen up to top-K\\% from each metric. As illustrated in Figure \\ref{fig:top_k}, the higher the restoration gap of the plan is, the poorer the performance is, which implies that the restoration gap captures the quality of the plan well compared to other metrics.", "candidate_comments": ["We first investigate whether the restoration gap can effectively identify low-quality plans. In order to do this, we compare our metric with realism score \\cite{kynkaanniemi2019improved} and rarity score \\cite{han2022rarity}, both of which are designed to assess the quality of individual generated samples by examining the discrepancy between the generated sample and the real data manifold in feature space. As Figure \\ref{fig:top_k} illustrates, the restoration gap demonstrates a stronger correlation between score values of generated plans and planning performance compared to other metrics. This reveals a considerable relationship between high restoration gap and poor planning performance, although the realism score and rarity score have limitations as they require real samples for precise real manifold estimation, in contrast to our restoration gap."]}
{"annotations": [{"id": 86985214, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986485, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86986495, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-2cf1ada74b2f65c2.jsonl", "id_parag": "2310.19427-2cf1ada74b2f65c2_3", "final_parag": "\\caption{Diffuser with RGG and Diffuser with RGG+ outperform all baselines. We report the mean and the standard error over 1000 planning seeds.}", "candidate_comments": ["\\caption{The Diffuser with RGG (DRGG) and Diffuser with RGG+ (DRGG+) performs better than the best prior model-free and model-based algorithms. Results for Diffuser, DRGG and DRGG+ correspond to the mean and standard error over 1000 planning seeds.}"]}
{"annotations": [{"id": 86724029, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}, {"id": 86730151, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86730505, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-2cf1ada74b2f65c2.jsonl", "id_parag": "2310.19427-2cf1ada74b2f65c2_4", "final_parag": "Maze2D environments \\cite{fu2020d4rl} involve a navigation task that requires an agent to exhibit long-horizon planning abilities to reach a target goal location. Maze2D environments consist of two tasks: a single-task where the goal location is fixed, and a multi-task which we refer to as Multi2D where the goal location is randomized at the beginning of every episode. We compare our methods with the model-free offline reinforcement learning algorithms CQL \\cite{kumar2020conservative} and IQL \\cite{kostrikov2021offline}; conventional trajectory optimizer MPPI \\cite{williams2015model}; and sequence modeling approach Diffuser \\cite{janner2022planning}. As shown in Table \\ref{maze2d-table}, RGG improves the planning performance of Diffuser in 5 out of 6 tasks, with notable improvements in the Maze2D-Large environments where the complexity of the obstacle maps is higher than in U-Maze or Medium layouts, leading to a higher occurrence of infeasible plans. RGG+ performs on par with or better than RGG. In contrast, model-free algorithms fail to reliably achieve the goal, as Maze2D environments require hundreds of steps to arrive at the goal location.", "candidate_comments": ["Maze2D environments \\cite{fu2020d4rl} involve a navigation task that requires an agent to exhibit long-horizon planning abilities to reach a target goal location, which offers a reward of 1. No reward shaping is provided at any other location. In Maze2D environments, there are three distinct maze layouts: \"U-Maze\", \"Medium\", and \"Large\", each offering different levels of difficulty. In Maze2D environments, there are two tasks: a single-task, where the goal location is fixed, and a multi-task, which we refer to as Multi2D, where the goal location is randomized at the beginning of every episode.", "We compare with the model-free offline reinforcement learning algorithms CQL \\cite{kumar2020conservative} and IQL \\cite{kostrikov2021offline}; conventional trajectory optimizer MPPI \\cite{williams2015model}; and sequence modeling approach Diffuser \\cite{janner2022planning}. As shown in Table \\ref{maze2d-table}, RGG improves the planning performance of Diffuser in 5 out of 6 tasks, with notable improvements in the Maze2D-Large environments where the complexity of the obstacle maps is higher than in U-Maze or Medium layouts, leading to a higher occurrence of infeasible plans. RGG+ performs on par with or better than RGG. In contrast, model-free algorithms fail to reliably achieve the goal, as Maze2D environments may require hundreds of steps to arrive at the goal location."]}
{"annotations": [{"id": 86865038, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86900192, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86900652, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19427-2cf1ada74b2f65c2.jsonl", "id_parag": "2310.19427-2cf1ada74b2f65c2_5", "final_parag": "Gym-MuJoCo locomotion tasks \\cite{fu2020d4rl} are standard benchmarks in evaluating algorithms on heterogeneous data with varying quality. We compare our methods with the model-free algorithms CQL \\cite{kumar2020conservative} and IQL \\cite{kostrikov2021offline}; model-based algorithms MOPO \\cite{yu2020mopo}, MOReL \\cite{kidambi2020morel}, and MBOP \\cite{argenson2020model}; sequence modeling approach Decision Transformer (DT) \\cite{chen2021decision}, Trajectory Transformer (TT) \\cite{janner2021offline} and Diffuser \\cite{janner2022planning}; and pure imitation-based approach behavior-cloning (BC). As indicated in Table \\ref{locomotion-table}, our approach of refining Diffuser with RGG either matches or surpasses most of the offline RL baselines when considering the average score across various tasks. Additionally, it significantly enhances the performance of Diffuser, particularly in the \"Medium\" dataset. We attribute this improvement to the sub-optimal and exploratory nature of the policy that was used to generate the \"Medium\" dataset, which results in a challenging data distribution to learn the diffusion planner. Consequently, RGG clearly contributes to the enhancement of planning performance. However, RGG+ only brings about a marginal improvement over RGG. This might be because we adopt the strategy of \\cite{janner2022planning}, using a closed-loop controller and a shorter planning horizon in locomotion environments compared to Maze2D environments, thereby simplifying the learning process of the gap predictor.", "candidate_comments": ["We compare with the model-free algorithms CQL \\cite{kumar2020conservative} and IQL \\cite{kostrikov2021offline}; model-based algorithms MOPO \\cite{yu2020mopo}, MOReL \\cite{kidambi2020morel}, and MOPO \\cite{yu2020mopo}; sequence modeling approach Decision Transformer (DT) \\cite{chen2021decision}, Trajectory Transformer (TT) \\cite{janner2021offline} and Diffuser \\cite{janner2022planning}; and pure imitation-based approach behavior-cloning (BC). As indicated in Table \\ref{locomotion-table}, our approach of refining Diffuser with RGG either matches or surpasses most of the offline RL baselines. Additionally, it significantly enhances the performance of Diffuser, particularly in the \"Medium\" dataset. We attribute this improvement to the sub-optimal and exploratory nature of the policy that was used to generate the \"Medium\" dataset, which results in a challenging data distribution to learn diffusion planner. Consequently, RGG clearly contributes to the enhancement of planning performance. However, RGG+ only brings about a marginal improvement over RGG. This might be because we adopt the strategy of \\cite{janner2022planning}, using a closed-loop controller and a shorter planning horizon in locomotion environments compared to Maze2D environments, thereby simplifying the learning process of the gap predictor."]}
{"annotations": [{"id": 86986529, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86987128, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86989253, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-2cf1ada74b2f65c2.jsonl", "id_parag": "2310.19427-2cf1ada74b2f65c2_6", "final_parag": "The explainability of decision-making models is particularly important in control domains as they could potentially harm physical objects including humans \\cite{kim2021explaining, lee2023adaptive, beechey2023explaining, kim2023variational, kenny2023towards}. \nTraining the gap predictor enables the diffusion planner to have explainability. Diffusion planners often generate trajectories with unreliable transitions resulting in execution failures. Attribution maps from the gap predictor highlight such unreliable transitions by identifying the extent to which each transition contributes to the decision of the gap predictor. Specifically, in Maze2D, the attribution maps emphasize the transitions involving wall-crossing or abrupt directional changes, as illustrated in Figure \\ref{fig:attribution_maps}. In the unconditional block stacking task where the robot destroys the tower while stacking the last block, the tower-breaking transitions are highlighted. On the other hand, for successful trajectories on the second and third attribution maps, the attribution maps do not emphasize picking or stacking behaviors. Similarly, in the conditional block stacking task where the robot fails to stack the block, they spotlight the transitions of stacking behaviors.", "candidate_comments": ["In many cases, diffusion planners might produce trajectories that contain unreliable transitions. As illustrated in Figure \\ref{fig:attribution_maps}, even in Maze2D environments, there is a high incidence of infeasible trajectories leading to execution failures. In such cases, attribution maps from the gap predictor can provide valuable insight into the plans generated by diffusion planners. The red areas on the attribution map indicate regions that greatly affect the prediction of the restoration gap. Transitions involving wall-crossings or abrupt directional changes within the generated trajectory are particularly highlighted. These observations validate the necessity of our restoration gap measure, as it successfully identifies and quantifies the problematic areas within the generated trajectories, thereby providing a way for refining the generative process of diffusion planners to generate more reliable trajectories."]}
{"annotations": [{"id": 86735652, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86902289, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86902425, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-37b23b8117e60f06.jsonl", "id_parag": "2310.19427-37b23b8117e60f06_1", "final_parag": "We investigate how sensitive the proposed method is depending on the choice of input attribution methods. As described in Section \\ref{sec:attribution_map_regularization}, RGG+ utilizes the input attribution method as a regularizer to prevent adversarial restoration gap guidance. We apply the input attribution method to the gap predictor, where it quantifies the impact of each transition in the generated plan on the prediction of the restoration gap. To validate the robustness of RGG+ to different input attribution methods, we compare its performance while employing three attribution methods: Grad-CAM \\cite{selvaraju2017grad}, Sailency \\cite{simonyan2013deep}, and DeepLIFT \\cite{shrikumar2017learning}. The experimental results for Maze2D tasks are presented in Table \\ref{ablation_attribution}. RGG+ exhibits comparable performances across the three attribution methods, which implies that our proposed method is robust in terms of the choice of input attribution methods.", "candidate_comments": ["We investigate the effectiveness of the attribution method on RGG+ performance by comparing three different attribution methods: Grad-CAM \\cite{selvaraju2017grad}, Sailency \\cite{simonyan2013deep}, and DeepLIFT \\cite{shrikumar2017learning}. Table \\ref{ablation_attribution} presents the experimental results, demonstrating comparable performances among the three attribution methods."]}
{"annotations": [{"id": 86739492, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86836873, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86898685, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-37b23b8117e60f06.jsonl", "id_parag": "2310.19427-37b23b8117e60f06_2", "final_parag": "To further investigate the effect of choosing different $\\hat{t}$ values on the restoration gap, we compare the performance of plans which are chosen up to top-K\\% from the restoration gap when setting $\\hat{t}$ to $0.3$, $0.5$, and $0.7$, in addition to the initial setting of $0.9$. The results of this ablation study are illustrated in Figure \\ref{fig:top_k_ablation_hat_t}. The results clearly show that the restoration gap remains significantly correlated with planning performance, even with these varied $\\hat{t}$ values, when considering the top $10$\\% of results. Specifically, while the correlation is strongest with $\\hat{t}$ set to $0.9$, even with $\\hat{t}$ at $0.3$, $0.5$, or $0.7$, the restoration gap still demonstrates a stronger correlation with planning performance compared to the rarity score or the negative realism score. This indicates the robustness of the restoration gap as a metric for assessing the quality of generated plans. Moreover, these results support our Proposition \\ref{prop_1} that a larger $\\hat{t}$ is an effective choice when using the restoration gap.", "candidate_comments": ["To further investigate the effect of different $\\hat{t}$ values on the restoration gap, we conducted an ablation study with $\\hat{t}$ set to $0.3$, $0.5$, and $0.7$, in addition to the initial setting of 0.9. The results of this study are illustrated in Figure \\ref{fig:top_k_ablation_hat_t}. It is clear from the results that the restoration gap remains significantly correlated with planning performance even with these varied $\\hat{t}$ values, when we consider the top 10\\% of results. Specifically, while the correlation is strongest with $\\hat{t}$ set to 0.9, even with $\\hat{t}$ at 0.3, 0.5, or 0.7, the restoration gap still demonstrated a stronger correlation with planning performance than either the rarity score or the negative realism score. This indicates the robustness of the restoration gap as a metric for assessing the quality of generated plans. Moreover, it supports our Proposition \\ref{prop_1} that a larger $\\hat{t}$ is an effective choice when using the restoration gap."]}
{"annotations": [{"id": 86872887, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86899865, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86900482, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19427-37b23b8117e60f06.jsonl", "id_parag": "2310.19427-37b23b8117e60f06_3", "final_parag": "We investigate how effectively the restoration gap evaluates the quality of generated samples by comparing the histogram with various metrics in Figure \\ref{fig:histogram} and comparing the planning performance in Figure \\ref{fig:top_k}. To further explore how significant a guidance signal provided by the restoration gap is for refining a diffusion planner, we compare the planning performance of plans guided by various metrics including the rarity score, the negative realism score, and the restoration gap. We conduct additional experiments in the Maze2D Large and Multi2D Large environments, the results of which are presented in Table \\ref{comp_to_other_metrics}. Consistent with the observations in the previous experiments, Table \\ref{comp_to_other_metrics} clearly demonstrates that the restoration gap is a useful metric for control tasks. Unlike other metrics requiring expert data for training, our restoration gap works without such constraints, making it even more practical.", "candidate_comments": ["To directly compare the performance of plans guided by different metrics, we run additional experiments in the Maze2D Large and Multi2D Large environments. Table \\ref{comp_to_other_metrics} clearly shows that the restoration gap is a useful metric for control tasks. Unlike other metrics that need expert data for training, our restoration gap works without such constraints, making it even more practical."]}
{"annotations": [{"id": 86961853, "completed_by": {"id": 62471}, "result": ["Réécriture", "Différent"]}, {"id": 86964834, "completed_by": {"id": 140156}, "result": ["Réécriture", "Différent"]}, {"id": 86964906, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent"]}], "agreement": 100.0, "source_file": "2310.19427-37b23b8117e60f06.jsonl", "id_parag": "2310.19427-37b23b8117e60f06_4", "final_parag": "Maze2D environments \\cite{fu2020d4rl} involve a navigation task where an agent needs to plan for a long-horizon to navigate toward a distant target goal location. A reward is not provided except when the agent successfully reaches the target goal where it only gets a reward of 1. Maze2D environments consist of three distinct maze layouts: \"U-Maze\", \"Medium\", and \"Large\" each of which offers different levels of difficulty. In Maze2D environments, there are two tasks: a single-task, where the goal location is fixed, and a multi-task, which we refer to as Multi2D, where the goal location is randomly selected at the beginning of every episode. Details about Maze2D environments are summarized in Table \\ref{tab:maze2d_env_details}.", "candidate_comments": ["Maze2D environments \\cite{fu2020d4rl} involve a navigation task that requires an agent to exhibit long-horizon planning abilities to reach a target goal location, which offers a reward of 1. No reward shaping is provided at any other location. In Maze2D environments, there are three distinct maze layouts: \"U-Maze\", \"Medium\", and \"Large\", each offering different levels of difficulty. In Maze2D environments, there are two tasks: a single-task, where the goal location is fixed, and a multi-task, which we refer to as Multi2D, where the goal location is randomized at the beginning of every episode.", "The objective of the unconditional stacking task is to construct a tower of blocks with maximum height. In this task, the diffusion planner generates a plan $\\boldsymbol{\\tau} \\in \\mathbb{R}^{39 \\times 128}$ where $39$ and $128$ represents the state dimensionality and episode length, respectively. In the conditional stacking task, where the objective is to stack blocks in a specified order, a plan $\\boldsymbol{\\tau} \\in \\mathbb{R}^{ \\times 128}$ is generated in a similar manner. Details about Block Stacking environments are summarized in Table \\ref{tab:kuka_env_details}."]}
{"annotations": [{"id": 86739342, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}, {"id": 86899594, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86900066, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-37b23b8117e60f06.jsonl", "id_parag": "2310.19427-37b23b8117e60f06_5", "final_parag": "The objective of the unconditional stacking task is to construct a tower of blocks with the maximum possible height. In this task, the agent observes the state including the joint position of the robot, as well as the position and rotation of each block, and then commands the robot's desired joint position while performing the grasping action to pick up the blocks. In the conditional stacking task, where the objective is to stack blocks in a specified order, the agent observes the same state as in the unconditional stacking task, but it additionally observes the index of the block which indicates the order in which the blocks should be stacked.", "candidate_comments": ["The objective of the unconditional stacking task is to construct a tower of blocks with maximum height. In this task, the diffusion planner generates a plan $\\boldsymbol{\\tau} \\in \\mathbb{R}^{39 \\times 128}$ where $39$ and $128$ represents the state dimensionality and episode length, respectively. In the conditional stacking task, where the objective is to stack blocks in a specified order, a plan $\\boldsymbol{\\tau} \\in \\mathbb{R}^{ \\times 128}$ is generated in a similar manner. Details about Block Stacking environments are summarized in Table \\ref{tab:kuka_env_details}.", "A state for both tasks includes the joint position of the robot, as well as the position, rotation, and grasping command of each block. For the conditional stacking task, an additional state, the index of the block is included, which indicates the order in which the blocks should be stacked. At each timestep, the agent controls the robot by commanding it to move to desired position and grasp the block."]}
{"annotations": [{"id": 86722627, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}, {"id": 86731834, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent"]}, {"id": 86728985, "completed_by": {"id": 140156}, "result": ["Réécriture", "Différent"]}], "agreement": 77.77777777777777, "source_file": "2310.19427-37b23b8117e60f06.jsonl", "id_parag": "2310.19427-37b23b8117e60f06_6", "final_parag": "We employ the same diffusion model for both tasks, but in the conditional stacking task, we additionally utilize a value function to guide the diffusion planner in stacking blocks according to specified conditions. Details about Block Stacking environments are summarized in Table \\ref{tab:kuka_env_details}.", "candidate_comments": ["The objective of the unconditional stacking task is to construct a tower of blocks with maximum height. In this task, the diffusion planner generates a plan $\\boldsymbol{\\tau} \\in \\mathbb{R}^{39 \\times 128}$ where $39$ and $128$ represents the state dimensionality and episode length, respectively. In the conditional stacking task, where the objective is to stack blocks in a specified order, a plan $\\boldsymbol{\\tau} \\in \\mathbb{R}^{ \\times 128}$ is generated in a similar manner. Details about Block Stacking environments are summarized in Table \\ref{tab:kuka_env_details}.", "A state for both tasks includes the joint position of the robot, as well as the position, rotation, and grasping command of each block. For the conditional stacking task, an additional state, the index of the block is included, which indicates the order in which the blocks should be stacked. At each timestep, the agent controls the robot by commanding it to move to desired position and grasp the block."]}
{"annotations": [{"id": 86962821, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86973069, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973862, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-74ab501d05f67524.jsonl", "id_parag": "2310.19427-74ab501d05f67524_1", "final_parag": "Diffusion models have gained prominence as a notable class of generative models, characterizing the data generation process through iterative denoising procedure \\cite{sohl2015deep, ho2020denoising}. This denoising procedure can be viewed as a way to parameterize the gradients of the data distribution \\cite{song2019generative}, linking diffusion models to score matching \\cite{hyvarinen2005estimation} and energy-based models (EBMs) \\cite{lecun2006tutorial, du2019implicit, nijkamp2019learning, grathwohl2020learning}. Recently, diffusion models have been successfully applied to various control tasks \\cite{janner2022planning, urain2022se, ajay2022conditional, chi2023diffusion, liang2023adaptdiffuser}. In particular, Diffuser \\cite{janner2022planning} employs an unconditional diffusion model to generate trajectories consisting of state-action pairs. The approach includes training a separate model that predicts the cumulative rewards of noisy trajectory samples, which then guides the reverse diffusion process towards high-return trajectory samples in the inference phase, analogous to classifier-guided sampling \\cite{dhariwal2021diffusion}. Building upon this, Decision Diffuser \\cite{ajay2022conditional} extends the capabilities of Diffuser by adopting a conditional diffusion model with reward or constraint guidance to effectively satisfy constraints, compose skills, and maximize return. Meanwhile, AdaptDiffuser \\cite{liang2023adaptdiffuser} enhances generalization ability of the diffusion model to unseen tasks by selectively fine-tuning it with high-quality data, derived through the use of hand-designed reward functions and an inverse dynamics model. In contrast, in this work, we focus on evaluating the quality of individually generated samples and explore ways to enhance planning performance by utilizing guidance derived from these evaluations.", "candidate_comments": ["Diffusion models have gained prominence as a notable class of generative models, characterizing the data generation process through iterative denoising procedure \\cite{sohl2015deep, ho2020denoising}. This denoising procedure can be viewed as a way to parameterize the gradients of the data distribution \\cite{song2019generative}, linking diffusion models to score matching \\cite{hyvarinen2005estimation} and energy-based models (EBMs) \\cite{lecun2006tutorial, du2019implicit, nijkamp2019learning, grathwohl2020learning}. Recently, diffusion models have been successfully applied to various control tasks \\cite{janner2022planning, urain2022se, ajay2022conditional, chi2023diffusion}. In particular, Diffuser \\cite{janner2022planning} employs an unconditional diffusion model to generate trajectories consisting of state-action pairs. The approach includes training a separate model that predicts the cumulative rewards of noisy trajectory samples, which then guides the reverse diffusion process towards high-return trajectory samples in the inference phase, analogous to classifier-guided sampling \\cite{dhariwal2021diffusion}. Decision Diffuser \\cite{ajay2022conditional} extends the capabilities of Diffuser by adopting a conditional diffusion model with reward or constraint guidance to effectively satisfy constraints, compose skills, and maximize return. In contrast, in this work, we focus on evaluating the quality of individually generated samples and explore ways to enhance planning performance by utilizing guidance derived from these evaluations."]}
{"annotations": [{"id": 86982437, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86984172, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86984442, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-9e3a530ff0809127.jsonl", "id_parag": "2310.19427-9e3a530ff0809127_1", "final_parag": "We have presented a novel refining method that fixes infeasible transitions within the trajectory generated by the diffusion planner. This refining process is guided by a proposed metric, restoration gap, which quantifies the restorability of a given plan. Under specific regularity conditions, we prove that the restoration gap effectively identifies unreliable plans while ensuring a low error probability for both type \\rom{1} and type \\rom{2} errors. The experimental results, which include enhancement in quantitative planning performance and visualization of qualitative attribution maps, highlight the importance of the refinement method of the diffusion planner.", "candidate_comments": ["We have presented a novel refining method which fixes infeasible transitions within the trajectory generated by the diffusion planner. Infeasible transitions are automatically detected by the proposed metric, restoration gap, which quantifies restorability of the given plan. Restoration gap theoretically bounds the error probability of identifying unreliable plans given a threshold value.  The experimental results, which include enhancement in quantitative planning performance and visualization of qualitative attribution maps, highlight the importance of refinement method of the diffusion planner. \nDetails regarding the limitations of this work can be found in Appendix \\ref{sec:limitations}."]}
{"annotations": [{"id": 86985249, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985454, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86985577, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-9e3a530ff0809127.jsonl", "id_parag": "2310.19427-9e3a530ff0809127_2", "final_parag": "While the restoration gap guidance effectively enhances the feasibility of plans and consistently improves the planning performance of diffusion models, our method is limited in situations where an offline dataset is provided. Training the diffusion model often requires transition data that uniformly covers the state-action space, the collection of which is a nontrivial and time-consuming task.", "candidate_comments": ["While the restoration gap guidance effectively enhances the feasibility of plans and consistently improves the planning performance of diffusion models, our method is limited in situations where an offline dataset is provided. Training the diffusion model often requires transition data that uniformly covers the state-action space, which is a nontrivial and time-consuming task."]}
{"annotations": [{"id": 86739397, "completed_by": {"id": 70661}, "result": ["Différent", "Réécriture"]}, {"id": 86837153, "completed_by": {"id": 62306}, "result": ["Différent", "Réécriture"]}, {"id": 86909888, "completed_by": {"id": 140156}, "result": ["Différent", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-9e3a530ff0809127.jsonl", "id_parag": "2310.19427-9e3a530ff0809127_3", "final_parag": "Our analysis of the effectiveness of the restoration gap is currently confined to a relatively simple task, Maze2D (see Figure \\ref{fig:histogram}), where we explicitly define normal and artifact plans. The choice of Maze2D is motivated by its suitability for identifying violations of prior knowledge, such as feasible plans not passing through walls. However, as future work, it would be worthwhile to explore the efficacy of restoration gap in more complex tasks, such as the block stacking task.", "candidate_comments": ["While the restoration gap guidance effectively enhances the feasibility of plans and consistently improves the planning performance of diffusion models, our method is limited in situations where an offline dataset is provided. Training the diffusion model often requires transition data that uniformly covers the state-action space, which is a nontrivial and time-consuming task.", "Furthermore, our analysis of proposed metric restoration gap is currently limited to a relatively simple task, Maze2D (see Figure \\ref{fig:histogram}). The choice of Maze2D is motivated by its suitability for identifying violations of the prior knowledge, such as plans should not pass through walls. However, as part of future work, it is appealing to explore the effectiveness of restoration gap in more complex tasks, such as block stacking task."]}
{"annotations": [{"id": 86738046, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86838524, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86898151, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19427-aa54d29279188aaf.jsonl", "id_parag": "2310.19427-aa54d29279188aaf_1", "final_parag": "\\let\\thefootnote\\relax \\makeatletter\\def\\Hy@Warning#1{}\\makeatother \\footnotetext{ Codes are available at \\href{https://github.com/leekwoon/rgg}{\\url{https://github.com/leekwoon/rgg}}. \n}", "candidate_comments": ["\\blfootnote{Codes are available at \\href{https://github.com/leekwoon/rgg}{\\url{https://github.com/leekwoon/rgg}}.}"]}
{"annotations": [{"id": 86739564, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86958784, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86961097, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-aa54d29279188aaf.jsonl", "id_parag": "2310.19427-aa54d29279188aaf_2", "final_parag": "We first propose a novel metric called \\textit{restoration gap} that can automatically detect whether generated plans are feasible or not. We theoretically analyze that it could detect artifacts with bounded error probabilities under regularity conditions. The restoration gap directly evaluates the quality of generated plans by measuring their restorability through diffusion models in which plans are exposed to a certain degree of noise, as illustrated in Figure \\ref{fig:restoration_gap_example}. A restoration gap is estimated by a function approximator which we name a \\textit{gap predictor}.", "candidate_comments": ["We first propose a novel metric called \\textit{restoration gap} that can directly evaluate the quality of generated plans by measuring their restorability when they are exposed to a certain degree of noise, as illustrated in Figure \\ref{fig:restoration_gap_example}."]}
{"annotations": [{"id": 86904475, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972843, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86972866, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19427-aa54d29279188aaf.jsonl", "id_parag": "2310.19427-aa54d29279188aaf_3", "final_parag": "This paper presents an orthogonal approach aimed at enhancing the plan quality of the diffusion model.", "candidate_comments": ["Next, we train a \\textit{gap predictor} using synthetic diffused data generated by the learned diffusion model."]}
{"annotations": [{"id": 86983004, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986254, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86986343, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19427-b384fbfb3b9e8bc7.jsonl", "id_parag": "2310.19427-b384fbfb3b9e8bc7_1", "final_parag": "The solution of a forward SDE is a time-varying random variable $\\boldsymbol{\\tau}^t$. \nUsing the reparameterization trick \\cite{kingma2013auto}, it is achieved by sampling a random noise $\\boldsymbol{\\epsilon}$ from a standard Gaussian distribution which is scaled by the target standard deviation $\\sigma_t$ and shifted by the target mean: where $\\alpha_t: [0,1] \\rightarrow[0,1]$ denotes a scalar function indicating the magnitude of the noiseless data $\\boldsymbol{\\tau}^0$, and ${\\sigma}_t: [0, 1] \\rightarrow [0, \\infty)$ denotes a scalar function that determines the size of the noise $\\boldsymbol{\\epsilon}$. Depending on perturbation strategies for $\\alpha_t$ and ${\\sigma}_t$, two types of SDEs are commonly considered: the Variance Exploding SDE (VE-SDE) has $\\alpha_t=1$ for all $t$; whereas the Variance Preserving (VP) SDE satisfies $\\alpha_t^2 + {\\sigma}_t^2 = 1$ for all $t$. Both VE and VP SDE change the data distribution to random Gaussian noise as $t$ moves from $0$ to $1$. In this work, we describe diffusion probabilistic models within the continuous-time framework using VE-SDE to simplify notation, as VE/VP SDEs are mathematically equivalent under scale translations \\cite{song2020score}.", "candidate_comments": ["The solution of a forward SDE is a time-varying random variable which we denote as $\\boldsymbol{\\tau}^t$, where $\\boldsymbol{\\tau}^0$ represents a noiseless trajectory data and $\\boldsymbol{\\tau}^t$ is distributed as a Gaussian distribution:"]}
{"annotations": [{"id": 86980980, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986125, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86985959, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19432-66d63f92c0a6e9d5.jsonl", "id_parag": "2310.19432-66d63f92c0a6e9d5_1", "final_parag": "However, for complex nonlinear models such as DNNs, it is problematic to approximate them with the first order Taylor expansion since it is difficult to find the root point and higher-order terms are ignored. Therefore, DTD approximates a layer-to-layer mapping with first order Taylor expansion, which enables one to decompose layer-wise mapping and compute the contributions of the layer-wise inputs which come from the outputs of previous layers.", "candidate_comments": ["Under the assumption of using a ReLU activation function, a function to approximate with the first order Taylor expansion is linear in the positive input space, which enables higher-order terms to be ignored. In addition, in the linear function, it is easier to find the root point $\\tilde{x}$."]}
{"annotations": [{"id": 86863786, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86900142, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86900568, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19449-b4e57ed2f529b85f.jsonl", "id_parag": "2310.19449-b4e57ed2f529b85f_1", "final_parag": "To address these requirements and enable fault injection as the default component in a machine learning development cycle, we introduce a novel fault injection framework called PyTorchALFI (Application Level Fault Injection for PyTorch) based on PyTorchFI. PyTorchALFI provides an efficient way to define randomly generated and reusable sets of faults to inject into PyTorch models, defines complex test scenarios, enhances data sets, and generates test KPIs while tightly coupling fault-free, faulty, and modified NN. In this paper, we provide details about the definition of test scenarios, software architecture, and several examples of how to use the new framework to apply iterative changes in fault location and number, compare different model modifications, and analyze test results.", "candidate_comments": ["Transient or permanent faults in hardware can modify the output of Neural Networks (NN) so that the results become incorrect. Modern NN possesses an inherent redundancy that can intercept several of these faults. To establish a safety case, it is necessary to quantify those faults that are not intercepted and appear as silent data errors (SDE) in incorrect NN predictions.  To study the effects of hardware (HW) faults on software (SW) in general and NN models in particular, several fault injection (FI) methods have been established in recent years. Current FI methods focus on the methodology of injecting faults but often fail to account for large-scale FI tests where many fault locations based on a particular fault model need to be analyzed in a short time; tests are distributed within a team, and results need to be concise, repeatable, and comparable. \nTo address these requirements and enable fault injection as the default component in a machine learning development cycle, we introduce a novel fault injection framework called PyTorchALFI (Application Level Fault Injection for PyTorch*) based on PyTorchFI. PyTorchALFI provides an efficient way to define randomly generated and reusable sets of faults to inject into PyTorch models, defines complex test scenarios, enhances data sets, and generates test KPI while tightly coupling fault-free, faulty, and modified NN. In this paper, we provide details about the definition of test scenarios, software architecture, and several examples of how to use the new framework to apply iterative changes in fault location and number, compare different model modifications, and analyze test results."]}
{"annotations": [{"id": 86952556, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86959090, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86959164, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_1", "final_parag": "\\begin{abstract} We propose principled Gaussian processes (GPs) for modeling functions defined over the edge set of a simplicial 2-complex, a structure similar to a graph in which edges may form triangular faces. \nThis approach is intended for learning flow-type data on networks where edge flows can be characterized by the discrete divergence and curl. \nDrawing upon the Hodge decomposition, we first develop classes of divergence-free and curl-free edge GPs, suitable for various applications. We then combine them to create \\emph{Hodge-compositional edge GPs} that are expressive enough to represent any edge function. These GPs facilitate direct and independent learning for the different Hodge components of edge functions, enabling us to capture their relevance during hyperparameter optimization. \nTo highlight their practical potential, we apply them for flow data inference in currency exchange, ocean flows and water supply networks, comparing them to alternative models.", "candidate_comments": ["We investigate principled approaches to define Gaussian processes (GPs) on the edge set of a simplicial 2-complex---a structure like graph in which edges may form triangular faces---for flow-type data on networks. \nEdge flows can be characterized by the discrete concepts of divergence and curl, measuring how they diverge at nodes and circulate along triangular faces. \nIn this paper, we first develop classes of divergence-free and curl-free edge GPs.  We then combine them---taking inspiration from the Hodge decomposition of edge functions---to create a more expressive family of \\emph{Hodge-compositional edge GPs} that can represent any edge functions. \nThis yields a principled prior that enables automatic relevance determination of different parts of the Hodge decomposition. \nWe validate its effectiveness in inferring flow data in currency exchange, ocean flows and water supply networks, comparing it to various alternatives."]}
{"annotations": [{"id": 86961940, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86981508, "completed_by": {"id": 140156}, "result": ["Différent", "Différent"]}, {"id": 86981284, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_2", "final_parag": "Our main contribution lies in the proposal of \\emph{Hodge-compositional edge GPs}. \nWe build them as combinations of three GPs, each modeling a specific part of the Hodge decomposition of an edge function, namely the gradient, curl and harmonic parts. \nWith a focus on the \\Matern~family, we show that each of them can be linked to a SPDE, extending the framework used by \\citet{borovitskiy2020matern,borovitskiyMatErnGaussian2021,borovitskiyIsotropicGaussianProcesses2023}. \nCompared to a direct extension of graph GPs, they enable separate learning of the different Hodge components, which allows us to capture the practical behavior of edge flows. \nWe also demonstrate their practical potential in edge-based learning tasks in foreign currency exchange markets, ocean flow analysis and water supply networks.", "candidate_comments": ["We start with deriving edge GPs from SPDEs on the edge space of SCs as a direct extension of graph counterparts.", "They can be built based on the Hodge Laplacian, encoding how edges are connected via nodes and triangular faces, which is a linear operator on edge functions, analogous to the graph Laplacian operating on function over nodes. "]}
{"annotations": [{"id": 86962443, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973141, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974444, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_3", "final_parag": "\\subsection{Div-free and Curl-free Edge GPs} The edge GPs in~\\Cref{subsec:non-hodge-gp} define distributions over all edge functions. \nAs opposed to this, here we seek to define GPs on the classes of divergence-free and curl-free edge functions. \nWe start with defining the appropriate notions of discrete derivatives, expressed in terms of the incidence matrices.", "candidate_comments": ["They can be measured in terms of notions of discrete derivatives, which are discrete analogs of $\\textGrad, \\textCurl$ and $\\textDiv$ operators in calculus and related to the incidence matrices."]}
{"annotations": [{"id": 86960867, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86979646, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent"]}, {"id": 86985192, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}], "agreement": 77.77777777777777, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_4", "final_parag": "\\begin{proposition} Let $\\vf_G$ and $\\vf_C $ be the gradient and curl Gaussian processes, respectively. Then, $\\textCurl\\,\\vf_G = \\vzero$ and $\\textDiv\\,\\vf_C=\\vzero$ with probability one. Moreover, a harmonic Gaussian process $\\vf_H$ follows $\\textCurl\\,\\vf_H = \\vzero$ and $\\textDiv\\,\\vf_H=\\vzero$ with probability one. \\end{proposition} See proof in \\cref{proof:prop-gp-samples}. \nThese Hodge GPs provide more targeted priors for special edge functions which are either div- or curl-free, capable of capturing these key properties. \nIn the case of \\Matern~kernels, we set [EQUATION] for $\\Box\\in\\{H,G,C\\}$, where $\\sigma_\\Box^2$ controls the variance we assign to the function in the subspace, and $\\nu_\\Box, \\kappa_\\Box$ are the regular \\Matern~parameters, as illustrated in \\cref{fig:spectral_kernel_illustration} (left). Note that since $\\mLambda_H=\\vzero$, we consider a scaling function for $\\mK_H$ as $\\Psi_H(\\vzero) = \\sigma_H^2$. These Hodge GPs can be derived from SPDEs on edges as well. \n\\begin{proposition} Given a scaled curl white noise $\\vw_C\\sim\\gN(\\vzero,\\mW_C)$ where $\\mW_C = \\sigma_C^2\\mU_C\\mU_C^\\top$, consider the following SPDE on edges: [EQUATION] with differential operators [EQUATION] The respective solutions give the curl edge GPs with \\Matern~kernel in \\cref{eq.grad-curl-matern-kernel} and diffusion kernel [EQUATION] Likewise, we can derive the gradient \\Matern~and diffusion GPs from the SPDEs as \\cref{eq.grad-curl-spde} but with operators $\\Phi_G(\\mL_{\\rmd})$ and a scaled gradient white noise. \\end{proposition} See proof in \\cref{app:derivation-grad-curl-gps}. We can draw the intuition of SPDE in \\cref{eq.grad-curl-spde} from the continuous analogy. In the case of $\\mL_{\\rmu}\\vf_C=\\vw_C$, the equation $\\textCurl^* \\textCurl \\, f_1(\\vx) = w_1(\\vx)$ is a stochastic vector Laplace's equation of a div-free (solenoidal) vector field, where $w_1(\\vx)$ the curl adjoint of some vector potential. In physics, this describes the static magnetic field from a magnetic vector potential, as well as an incompressible fluid.", "candidate_comments": ["\\begin{proposition}      Let $\\vf_G$ and $\\vf_C $ be the gradient and curl Gaussian processes in \\cref{eq.grad-curl-gps}, respectively.      Then, their samples are curl-free and divergence-free, respectively. Moreover, samples of a harmonic Gaussian process are both div- and curl-free. \n\\end{proposition}", "While they are of particular interest for edge functions that are in certain subspace, general edge functions do not always have such properties. "]}
{"annotations": [{"id": 86960527, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972853, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973820, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_5", "final_parag": "In the spatial/edge domain, this is related to separating the down and up adjacencies encoded in the SPDE operators $\\Phi(\\cdot)$. \nFrom an eigen-spectrum perspective, the eigenvalues $\\Psi_\\Box$ of HC GP's kernels associated to the three Hodge subspaces have individual parameters. \nThis enables capturing the different Hodge components of edge functions, as well as their relevance during hyperparameter optimization, further allowing us to recover the Hodge components in predictions, which we detail in \\cref{app-subsec:posterior-hodge-components}. \nAnother implication is that, unlike for GPs from~\\Cref{subsec:divcurl_free_gps}, we do not require specific knowledge about the div or curl of the underlying function.", "candidate_comments": ["Also, by considering $\\mK_G$ in the posterior mean instead of $\\mK_1$, we can obtain the gradient part of predictions for analysis purpose, and likewise for the other parts.  Also, we can obtain the Hodge components of the prediction from the optimized corresponding kernel in regression problems. "]}
{"annotations": [{"id": 86904886, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86973029, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974178, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_6", "final_parag": "Compared to the graph (node) diffusion converging ($t\\to\\infty$) to the state that is constant on all nodes as long as the graph is connected, the harmonic state of the edge diffusion can be non-constant, lying in the span of $\\mU_H$.", "candidate_comments": ["While both edge and graph (node) diffusion processes converge towards harmonic states as $t\\to\\infty$, the harmonic state of the former can be non-constant, as opposed to that of the latter, as long as the graph is connected, as shown in \\cref{fig:diffusion_demonstration}.  "]}
{"annotations": [{"id": 86731708, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86732142, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86766223, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_7", "final_parag": "\\paragraph{Complexity} The kernels of HC edge GPs can be constructed in a scalable way by considering the $l$ largest eigenvalues with off-the-shelf eigen-solvers, e.g., Lanczos algorithm. See \\cref{app-subsec:complexity} for more details on the complexity of HC GPs.", "candidate_comments": ["\\paragraph{Efficiency} HC GPs have the same complexity as non-HC ones. \nThe \\Matern~and diffusion kernels can be trained in a scalable way.  Due to their decreasing eigenvalues, we can consider the $l$ largest eigenvalues with off-the-shelf eigen-solvers, e.g., Lanczos algorithm.  Other computational techniques applicable for graph GPs in \\citet{borovitskiyMatErnGaussian2021,nikitinNonseparableSpatiotemporalGraph2022} can be adopted as well. "]}
{"annotations": [{"id": 86985677, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86987112, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86990147, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_8", "final_parag": "The proof follows from (i) derivatives preserving Gaussianity, and (ii) $\\mL_0$ and $\\mL_{\\rmd}$ having the same nonzero eigenvalues. \nWe can also obtain a curl edge GP from a GP on triangles likewise. In turn, for an edge GP, its div is a node GP and its curl is a GP on triangles. \nWe refer to \\cref{app:interaction-node-edge} for the proof and more details.", "candidate_comments": ["Thus, we can define edge GPs from node and/or triangle GPs because the gradient and curl parts of edge functions come from the (co)derivative of some node and triangle functions as in \\cref{eq.hodge-composition-expression}."]}
{"annotations": [{"id": 86984791, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86986437, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86989292, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_9", "final_parag": "See proof in \\cref{app:alternative_gp}. This alternative HC GP incorporates the Hodge theorem prior in a way that directly relates the node potential and the triangle function. It can be applicable when GP priors of node or triangle functions are more discernible.", "candidate_comments": ["They are more applicable when the underlying physical relationships exist between the corresponding functions and the GP priors on the original simplices are easier to construct."]}
{"annotations": [{"id": 86910273, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86958476, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86965592, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_10", "final_parag": "We then consider the edge-based ocean flow learning following the setup in \\citet{chenHelmholtzianEigenmapTopological2021}. The flow velocity fields in the ocean were converted using the linear integration approximation to edge flows within a $\\textSC$ whose nodes are 1500 buoys sampled from North Pacific ocean drifter records in 2010-2019 \\citep{lumpkin2019global}. \nWe apply both non-HC and HC GP models to predict the converted edge flows. \nGiven the large number of edges ($\\sim$20k), we consider a truncated approximation of kernels with eigenpairs associated with the 500 largest eigenvalues \\citep{knyazev2001toward}.", "candidate_comments": ["The Hodge Laplacian based approaches were applied to ocean flow velocity learning by \\citet{chenHelmholtzianEigenmapTopological2021} with better efficiency compared to the traditional one \\citep{berry2020spectral}.  They constructed a $\\rm{SC}_2$ of 1500 buoys sampled from North Pacific ocean flow records in 2010-2019 \\citep{lumpkin2019global} and converted the velocity fields to edge flows using the linear integration approximation \\citep[Appendix H]{chenHelmholtzianEigenmapTopological2021}. "]}
{"annotations": [{"id": 86904122, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}, {"id": 86972497, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86972850, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent"]}], "agreement": 77.77777777777777, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_11", "final_parag": "We introduced Hodge-compositional (HC) Gaussian processes (GPs) for modeling functions on the edges of simplicial 2-complexes. These HC GPs are constructed by combining three individual GPs, each designed to capture the gradient, curl and harmonic components of the Hodge decomposition of edge functions. This allows them to independently learn each component, making them more expressive and interpretable when compared to various alternatives. \nThey can also be constructed by leveraging the physical interactions between functions on nodes, edges and triangles. \nWe demonstrated their practical potential in learning real-world flow data.", "candidate_comments": ["We introduced Hodge-compositional (HC) Gaussian processes, a principled framework for modeling functions over the edges of simplicial 2-complexes. \nWe build them as the combination of three GPs---each with individual kernel hyperparameters---on the gradient, curl and harmonic parts of edge functions given by the Hodge decomposition. \nEach of the three can be linked to their respective SPDE representation on edges.  Compared to the non-HC edge GPs, directly extended from graph GPs, HC GPs benefit from the separate learning of each Hodge part, having better expressivity and interpretablity. \nWe showed the practical potentials of the HC GPs in statistical learning of flow-type data related to currency exchange rates, ocean currents and water supply networks. ", "We also showed they admit an alternative construction that exploits the interactions between functions on nodes, edges and triangles.  We hope our work inspires further advancements in the field of non-Euclidean GPs. \nThis principled approach into statistical learning in complex network domains."]}
{"annotations": [{"id": 86984626, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86986469, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86986628, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_12", "final_parag": "Here we discuss the div and curl properties of samples of gradient and curl GPs in \\cref{eq.grad-curl-gps}, which completes the proof of \\cref{prop:samples-of-grad-curl-gps} . \\begin{proposition} Consider the gradient and curl GPs [EQUATION] where the gradient kernel and the curl kernel are [EQUATION] Their prior samples are, respectively, curl-free and div-free. \\end{proposition}", "candidate_comments": ["\\begin{proposition}      Let $\\vf_G$ and $\\vf_C $ be the gradient and curl Gaussian processes in \\cref{eq.grad-curl-gps}, respectively.      Then, their samples are curl-free and divergence-free, respectively. Moreover, samples of a harmonic Gaussian process are both div- and curl-free. \n\\end{proposition}"]}
{"annotations": [{"id": 86952759, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86959915, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86960757, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_13", "final_parag": "\\begin{proof} We focus on the case of gradient GPs. \nFirst, we can decompose the gradient kernel in terms of $\\mU_1 = [\\mU_H\\,\\,\\mU_G\\,\\,\\mU_C]$ as [EQUATION] From a vector $\\vv=(v_1,\\dots,v_{N_1})^\\top$ of variables following independent normal distribution, we can draw a random sample of gradient function as [EQUATION] where $\\diag([\\va,\\vb,\\vc])$ is the diagonal matrix with $(\\va,\\vb,\\vc)^\\top$ on its diagonal.", "candidate_comments": ["An alternative proof can follow by studying the curl of the gradient GP which is another GP on triangles "]}
{"annotations": [{"id": 86962969, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}, {"id": 86985180, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86985215, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_14", "final_parag": "An alternative proof can follow by studying the curl of the gradient GP which is another GP on triangles as given later by \\cref{app-prop:div-curl-gp}. \nThe kernel $\\mB_2^\\top\\mK_G\\mB_2$ is zero, due to the orthogonality $\\mB_2^\\top\\mU_G = \\vzero$. Thus, the curl of a gradient GP is a zero GP on triangles, as well as its samples. Similarly, one can show the div of a curl GP is a zero GP on nodes, thus, its samples are zero. \\end{remark} \\end{proof}", "candidate_comments": ["An alternative proof can follow by studying the curl of the gradient GP which is another GP on triangles ", "but the kernel $\\mB_2^\\top\\mK_G\\mB_2$ is zero, due to the orthogonality $\\mB_2^\\top\\mU_G = \\vzero$.  Thus, the curl of a gradient GP is a zero GP on triangles, as well as its samples. Similarly, one can show the div of a curl GP is a zero GP on nodes, thus, its samples are zero. "]}
{"annotations": [{"id": 86899650, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86901854, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86909969, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19450-9b1ae365a8285313.jsonl", "id_parag": "2310.19450-9b1ae365a8285313_15", "final_parag": "\\begin{proposition}[Div and Curl of edge GPs] Let $\\vf_1\\sim \\gG\\gP(\\vzero,\\mK)$ be a Hodge-compositional edge Gaussian process in \\cref{def:hodge-compositional-gp}. Then, its divergence and curl are Gaussian processes on nodes and triangles, respectively, as follows [EQUATION] \\end{proposition}", "candidate_comments": ["\\begin{proposition}   Let $\\vf\\sim \\gG\\gP(\\vm,\\mK)$ be an edge Gaussian process. Then, its divergence and curl are also Gaussian processes with   \\begin{subequations}   \\end{subequations}   where $\\vm_G$ and $\\vm_C$ are the gradient and curl components of $\\vm$, respectively. \n\\end{proposition}"]}
{"annotations": [{"id": 86737253, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}, {"id": 86902219, "completed_by": {"id": 62306}, "result": ["Réécriture", "Différent"]}, {"id": 86903741, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}], "agreement": 77.77777777777777, "source_file": "2310.19455-e1425d772d0e14e8.jsonl", "id_parag": "2310.19455-e1425d772d0e14e8_1", "final_parag": "\\switch{ \\fancyfoot[R]{\\scriptsize{Copyright \\textcopyright\\ 2024 by SIAM\\\\ Unauthorized reproduction of this article is prohibited}} }{}", "candidate_comments": ["\\fancyfoot[R]{\\scriptsize{Copyright \\textcopyright\\ 20XX\\\\ Copyright for this paper is retained by authors}}", "\\fancyfoot[R]{\\scriptsize{Copyright \\textcopyright\\ 20XX\\\\ Copyright retained by principal author's organization}}"]}
{"annotations": [{"id": 86726971, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86731971, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86728915, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19455-e1425d772d0e14e8.jsonl", "id_parag": "2310.19455-e1425d772d0e14e8_2", "final_parag": "Interestingly, the \\myproblem{popular common independent set} problem which asks for a common independent set that is popular in the set of all common independent sets (of all sizes) in the matroid intersection can be reduced to the \\myproblem{popular common base} problem (see Section~\\ref{sec:colorful}). Therefore, the following fact is obtained as a corollary to Theorem~\\ref{thm:pop-largest}. \n\\begin{corollary}", "candidate_comments": ["In general, the matroid intersection need not admit common bases, and in such a case, an alternative is a largest common independent set  that is popular among all largest common independent sets. This problem can be easily reduced to the popular common base problem  (see Appendix~\\ref{sec:discussion})."]}
{"annotations": [{"id": 86904144, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86952891, "completed_by": {"id": 140156}, "result": ["Réécriture", "Différent"]}, {"id": 86960693, "completed_by": {"id": 70661}, "result": ["Réécriture", "Différent"]}], "agreement": 77.77777777777777, "source_file": "2310.19455-e1425d772d0e14e8.jsonl", "id_parag": "2310.19455-e1425d772d0e14e8_3", "final_parag": "All of the following problems fall in the framework of a popular common base (or common independent set) in the intersection of a partition matroid with another matroid: \\begin{enumerate} \\item Popular matchings~\\cite{AIKM07}. \n\\item Popular assignments~\\cite{KKMSS22}. \n\\item Popular branchings~\\cite{KKMSS20}. \n\\item Popular matchings with matroid constraints\\footnote{This problem asks for a popular many-to-one matching in a bipartite graph $G = (A \\cup B, E)$ where vertices in $A$ have weak rankings and the vertices that get matched to each $b \\in B$ must form an independent set in a matroid $M_b$.}~\\cite{Kam17}. \n\\end{enumerate}", "candidate_comments": ["(i)~popular matchings~\\cite{AIKM07}, (ii)~popular assignments~\\cite{KKMSS22}, (iii)~popular branchings~\\cite{KKMSS20}, and (iv)~popular matchings with matroid constraints\\footnote{This problem asks for a popular many-to-one matching in a bipartite graph $G = (A \\cup B, E)$ where vertices in $A$ have weak rankings and the vertices that get matched to each $b \\in B$ must form an independent set in a matroid $M_b$.}~\\cite{Kam17}.", "for both the \\myproblem{popular branching} problem in~\\cite{KKMSS20} and for the \\myproblem{popular assignment} problem in~\\cite{KKMSS22}."]}
{"annotations": [{"id": 86737851, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}, {"id": 86901656, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}, {"id": 86901958, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19455-e1425d772d0e14e8.jsonl", "id_parag": "2310.19455-e1425d772d0e14e8_4", "final_parag": "\\item $\\lev_{\\C}(e)=\\lev^*_{\\C}(v)$ and there is no element $e' \\in \\delta(v)$ such that $\\lev_{\\C}(e')=\\lev^*_{\\C}(v)$ and $e'\\succ_v e$;", "candidate_comments": ["\\item $\\lev_{\\C}(e)=\\lev^*_{\\C}(v)$ and $e\\succeq_v e'$ for every $e' \\in \\delta(v)$ such that $\\lev_{\\C}(e')=\\lev^*_{\\C}(v)$;", "\\item $\\lev_{\\C}(e)=\\lev^*_{\\C}(v)-1$ and $e\\succeq_v e'$ for every $e' \\in \\delta(v)$ such that $\\lev_{\\C}(e')=\\lev^*_{\\C}(v)-1$, and moreover, $e\\succ_v f$ for every $f \\in \\delta(v)$ with $\\lev_{\\C}(f)=\\lev^*_{\\C}(v)$."]}
{"annotations": [{"id": 86903926, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86982744, "completed_by": {"id": 126844}, "result": ["Différent", "Différent"]}, {"id": 86982775, "completed_by": {"id": 140156}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19455-e1425d772d0e14e8.jsonl", "id_parag": "2310.19455-e1425d772d0e14e8_5", "final_parag": "\\item $\\lev_{\\C}(e)=\\lev^*_{\\C}(v)-1$ and there is no element $e' \\in \\delta(v)$ such that $\\lev_{\\C}(e')=\\lev^*_{\\C}(v)-1$ and $e'\\succ_v e$, and moreover, $e\\succ_v f$ for every $f \\in \\delta(v)$ with $\\lev_{\\C}(f)=\\lev^*_{\\C}(v)$.", "candidate_comments": ["\\item $\\lev_{\\C}(e)=\\lev^*_{\\C}(v)$ and $e\\succeq_v e'$ for every $e' \\in \\delta(v)$ such that $\\lev_{\\C}(e')=\\lev^*_{\\C}(v)$;", "\\item $\\lev_{\\C}(e)=\\lev^*_{\\C}(v)-1$ and $e\\succeq_v e'$ for every $e' \\in \\delta(v)$ such that $\\lev_{\\C}(e')=\\lev^*_{\\C}(v)-1$, and moreover, $e\\succ_v f$ for every $f \\in \\delta(v)$ with $\\lev_{\\C}(f)=\\lev^*_{\\C}(v)$."]}
{"annotations": [{"id": 86736501, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86836852, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86898620, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19455-e1425d772d0e14e8.jsonl", "id_parag": "2310.19455-e1425d772d0e14e8_6", "final_parag": "\\end{itemize} In other words, $e \\in \\delta(v)$ belongs to $E(\\C)$ if either (i)~$e$ is a maximal element in $\\delta(v)$ with respect to $\\succ_v$ among those in $\\lev^*_{\\C}(v)$ or (ii)~$e$ is a maximal element in $\\delta(v)$ among those in $\\lev^*_{\\C}(v) - 1$ and $v$ strictly prefers $e$ to all elements in level $\\lev^*_{\\C}(v)$. \nFrom Lemma~\\ref{lem:duality}, we obtain the following useful characterization of popular arborescences.", "candidate_comments": ["\\item $\\lev_{\\C}(e)=\\lev^*_{\\C}(v)-1$ and $e\\succeq_v e'$ for every $e' \\in \\delta(v)$ such that $\\lev_{\\C}(e')=\\lev^*_{\\C}(v)-1$, and moreover, $e\\succ_v f$ for every $f \\in \\delta(v)$ with $\\lev_{\\C}(f)=\\lev^*_{\\C}(v)$."]}
{"annotations": [{"id": 86984683, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86985262, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985240, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19455-e1425d772d0e14e8.jsonl", "id_parag": "2310.19455-e1425d772d0e14e8_7", "final_parag": "\\subsection{Example~1.} This instance is similar to the one illustrateded in Section~\\ref{sec:intro}; the only difference is that now the edge $(r, d)$ is deleted. \nIn contrast to the case where $(r,d)$ exists, this instance admits a popular arborescence, which is found by Algorithm~\\ref{alg:pop-arb} as follows.", "candidate_comments": ["This instance is similar to the one described in Section~\\ref{sec:intro}; the only difference is that now the edge $(r, d)$ is deleted. \nIn contrast to the case where $(r,d)$ exists, this instance admits a popular arborescence, which is found by Algorithm~\\ref{alg:pop-arb} as follows. "]}
{"annotations": [{"id": 86732237, "completed_by": {"id": 70661}, "result": ["Réécriture", "Différent"]}, {"id": 86732831, "completed_by": {"id": 140156}, "result": ["Réécriture", "Différent"]}, {"id": 86766028, "completed_by": {"id": 62306}, "result": ["Réécriture", "Différent"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_1", "final_parag": "Constituency parsing is a fundamental yet unsolved natural language processing task. \nIn this paper, we explore the potential of recent large language models (LLMs) that have exhibited remarkable performance across various domains and tasks to tackle this task. \nWe employ three linearization strategies to transform output trees into symbol sequences, such that LLMs can solve constituency parsing by generating linearized trees. \nWe conduct experiments using a diverse range of LLMs, including ChatGPT, GPT-4, OPT, LLaMA, and Alpaca, comparing their performance against the state-of-the-art constituency parsers. Our experiments encompass zero-shot, few-shot, and full-training learning settings, and we evaluate the models on one in-domain and five out-of-domain test datasets. \nOur findings reveal insights into LLMs' performance, generalization abilities, and challenges in constituency parsing.", "candidate_comments": ["Constituency parsing is a fundamental yet unsolved natural language processing task with significant implications for various downstream applications. To better understand the potential of large language models (LLMs) in addressing this challenge, we investigate their impact on constituency parsing.  We adapt generative LLMs by transforming constituency trees into symbol sequences using various linearization strategies.  Our evaluation includes diverse LLMs, such as ChatGPT, GPT-4, LLaMA, and Alpaca, and compares their performance with state-of-the-art constituency parsers.  We conduct experiments under zero-shot, few-shot, and full-training learning settings on one in-domain test dataset and five out-of-domain test datasets to comprehensively assess the models.  Our findings reveal insights into LLMs' performance and generalization abilities in constituency parsing. \n\\xfbai{", "} , and we provide an error analysis to further understand the main challenges faced by LLMs-based models in this task."]}
{"annotations": [{"id": 86904683, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86986667, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86986743, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_2", "final_parag": "Constituency parsing is a fundamental natural language processing task which aims to predict the syntactic structure of a given sentence according to a phrase structure grammar~\\cite{10.5555/972470.972475,collins-1997-three,charniak-2000-maximum, petrov-klein-2007-improved}. \nAs shown in Figure~\\ref{fig:example}, given the sentence ``\\textit{Singapore is located in Asia.}'', a constituent parser produces its phrase structure tree automatically. \nThis task is not only appealing in computational linguistics studies but also serves as a foundation for improving various natural language processing tasks over the past decades, facilitating numerous downstream applications~\\cite{wang-etal-2018-tree,jiang-diesner-2019-constituency,xu-durrett-2019-neural}. \nDespite its importance, constituency parsing remains a challenging problem, with the best results on the news domain being 96\\%~\\cite{kitaev-etal-2019-multilingual}, while the performances on review and literature domains being only 84\\% and 86\\%, respectively~\\cite{yang-etal-2022-challenges}.", "candidate_comments": ["Constituency parsing is a fundamental natural language processing task which aims to predict the syntactic structure of a given sentence according to a phrase structure grammar~\\cite{10.5555/972470.972475,collins-1997-three}.  In addition to its own appeal in computational linguistics studies, constituency parsing has also served as a basis for improving a range of natural language processing tasks over the past decades and it facilitates a wide set of downstream applications~\\cite{wang-etal-2018-tree,jiang-diesner-2019-constituency,xu-durrett-2019-neural}. \nToday, constituency parsing is still not a solved problem, with the best results on the news domain being 96\\% while the best performances on biomedical and review domains being 86\\% and 88\\% respectively. "]}
{"annotations": [{"id": 86734562, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86837194, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86898201, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_3", "final_parag": "Recently, large language models (LLMs) trained on massive data have achieved remarkable performance on many NLP tasks, such as text classification~\\cite{schick2020exploiting, chen-etal-2022-adaprompt}, commonsense reasoning~\\cite{zhang2022task}, text summarization~\\cite{goyal2022news, chen2022unisumm}, machine translation~\\cite{jiao2023chatgpt} and dialogue systems~\\cite{qin2023chatgpt}, using prompting technologies. \nThe primary idea is to convert NLP tasks into a text-to-text problem, enabling knowledge transfer from pre-training tasks to downstream tasks.", "candidate_comments": ["Recently, large language models (LLMs) trained on massive data have achieved remarkable performance on many NLP tasks, such as text classification and commonsense reasoning using, prompting technologies~\\cite{Wei0SBIXCLZ22, chen-etal-2022-adaprompt,zhang2022automatic,zhang2023multimodal}. \nThe main idea is to convert NLP tasks into a text-to-text manner, which allows knowledge to transfer from pre-training tasks to downstream tasks. \nHowever, there has been few research utilizing LLMs for constituency parsing.  One primary reason is that the output of constituency parsing is represented in a tree structure, which is rarely found in textual data - the primary format where LLMs are trained. \nIt still remains an open-question whether LLMs can generalize well on constituency parsing."]}
{"annotations": [{"id": 86983905, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86986512, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86986693, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_4", "final_parag": "However, due to the fact that the output of constituency parsing is represented in a tree structure, which is rarely found in textual data where LLMs are trained, there has been limited research utilizing LLMs for constituency parsing. \nIt remains an open question whether LLMs can generalize well on constituency parsing.", "candidate_comments": ["Recently, large language models (LLMs) trained on massive data have achieved remarkable performance on many NLP tasks, such as text classification and commonsense reasoning using, prompting technologies~\\cite{Wei0SBIXCLZ22, chen-etal-2022-adaprompt,zhang2022automatic,zhang2023multimodal}. \nThe main idea is to convert NLP tasks into a text-to-text manner, which allows knowledge to transfer from pre-training tasks to downstream tasks. \nHowever, there has been few research utilizing LLMs for constituency parsing.  One primary reason is that the output of constituency parsing is represented in a tree structure, which is rarely found in textual data - the primary format where LLMs are trained. \nIt still remains an open-question whether LLMs can generalize well on constituency parsing.", "However, there has been limited research utilizing LLMs for constituency parsing. \nOne main reason is that the output of constituency parsing is represented in a tree structure, which is rarely found in textual data where LLMs are trained."]}
{"annotations": [{"id": 86961064, "completed_by": {"id": 62471}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86982465, "completed_by": {"id": 126844}, "result": ["Différent", "Réécriture", "Réécriture"]}, {"id": 86982248, "completed_by": {"id": 140156}, "result": ["Réécriture", "Différent", "Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_5", "final_parag": "To fill this gap, we take the first step to study the impact of LLMs on constituency parsing. \nIn order to enable LLMs to process constituency trees, we first transform constituency trees into sequences of symbols using various linearization strategies. \nAnd we empirically evaluate LLMs from the two most typical families, namely the GPTs~\\cite{BrownMRSKDNSSAA20,ouyang2022training,openai2023gpt} (including ChatGPT, GPT-4, etc) and LLaMAs~\\cite{touvron2023llama} (including Vicuna~\\cite{vicuna2023}, Alpaca~\\cite{alpaca}, etc), comparing them with the state-of-the-art constituency parsers~\\cite{kitaev-etal-2019-multilingual}. \nWe conduct experiments under multiple learning settings, with a zero-shot and a few-shot scenario to measure how much syntactic knowledge LLMs learned during pre-training, and a full-training setting to study to what extent LLMs boost the performance of constituency parsing.", "candidate_comments": ["However, there has been limited research utilizing LLMs for constituency parsing. \nOne main reason is that the output of constituency parsing is represented in a tree structure, which is rarely found in textual data where LLMs are trained.", "In this work, we take the first step to study the impact of LLMs on constituency parsing, investigating three main research questions:  1) {\\sl How well do LLMs (including public and non-public commercial LLMs) perform on constituency parsing?} 2) {\\sl Are LLMs still few-shot learners for constituency parsing?} 3) {\\sl What factors can influence the performance of LLMs.} In this work, we take the first step to study the impact of LLMs on constituency parsing, investigating three main research questions: 1. {\\sl How well do LLMs (including public and non-public commercial LLMs) perform on constituency parsing?} 2. {\\sl Are LLMs still few-shot learners for constituency parsing?} 3. {\\sl What factors can influence the performance of LLMs?}", "To adapt generatively-trained LLMs to constituency parsing, we first transform a constituency tree into a sequence of symbols using various linearization strategies. \nWe empirically choose a wide range of LLMs for evaluation, including ChatGPT~\\cite{BrownMRSKDNSSAA20,ouyang2022training}, GPT-4~\\cite{openai2023gpt}, LLaMA~\\cite{touvron2023llama} and Alpaca~\\cite{alpaca,wang2022self}, comparing their with state-of-the-art constituency parsers~\\cite{kitaev-etal-2019-multilingual}. \nWe conduct experiments under various learning settings, with a zero-shot and a few-shot setting to measure how much syntactic knowledge do LLMs learned during pre-training, a full-training setting to study to what extent do LLMs boost the performance of constituency parsing. \nWe evaluate the models on one in-domain test dataset and five out-of-domain test datasets to assess the model capability and cross-domain generalization abilities, respectively. \nThe experimental results show that 1) xxx; 2) xxx; 3) xxx; 4) xxx."]}
{"annotations": [{"id": 86979897, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86985372, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86989350, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_6", "final_parag": "We evaluate the state-of-the-art parsers and LLMs on one in-domain test dataset and five out-of-domain test datasets to measure their capability and cross-domain generalization abilities, respectively. \nExperimental results show that: 1) LLMs greatly boost the performance of sequence-based constituency parsing methods; 2) Fine-tuned LLM-based parsers can give competitive results compared to the state-of-the-art chart-based and transition-based parsers; 3) LLMs (except commercial ones) are not good few-shot learners for constituency parsing; 4) To our surprise, while LLMs improve the cross-domain generalization ability of sequence-based methods, LLM-based parsers are weaker than chat-based parsers in cross-domain parsing. \nFurthermore, we conduct an error analysis to study the main challenges of LLM-based models for constituency parsing, and our analyses show that: 1) LLM-based parsers suffer from hallucination and have limited ability to learn extremely long constituents; 2) The main challenge for LLM-based constituency parsing stems from invalid trees, which can be because that data of tree structures are seldomly presented in LLMs's pre-training. \nOur code and trained models are available at~\\url{https://github.com/goodbai-nlp/LLM-ConParsing}.", "candidate_comments": ["In this work, we take the first step to study the impact of LLMs on constituency parsing, investigating three main research questions:  1) {\\sl How well do LLMs (including public and non-public commercial LLMs) perform on constituency parsing?} 2) {\\sl Are LLMs still few-shot learners for constituency parsing?} 3) {\\sl What factors can influence the performance of LLMs.} In this work, we take the first step to study the impact of LLMs on constituency parsing, investigating three main research questions: 1. {\\sl How well do LLMs (including public and non-public commercial LLMs) perform on constituency parsing?} 2. {\\sl Are LLMs still few-shot learners for constituency parsing?} 3. {\\sl What factors can influence the performance of LLMs?}"]}
{"annotations": [{"id": 86721755, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86765416, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86826651, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_7", "final_parag": "This work belongs to the category of sequence-based methods. \nDifferent from existing methods, we solve constituency parsing with generative LLMs which have shown impressive performance in many NLP applications. \nTo our knowledge, we are the first to systematically study the impact of LLMs on constituency parsing.", "candidate_comments": ["However, sequence-based methods are susceptible to generating invalid trees when the opening and closing brackets do not correspond appropriately."]}
{"annotations": [{"id": 86738417, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86901340, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86901408, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_8", "final_parag": "\\textbf{Large Language Models}. \nLarge Language Models (LLMs)~\\cite{BrownMRSKDNSSAA20,chowdhery2022palm,touvron2023llama} have substantially influenced the field of Natural Language Processing (NLP). \nAs the pioneering work, ~\\citet{radford2019language} and~\\citet{BrownMRSKDNSSAA20} demonstrate the capability of language models to solve a task with minimal task supervision. \nThe following work shows that LLMs are adept at leveraging textual instructions to perform various tasks~\\cite{sanh2022multitask,webson-pavlick-2022-prompt,liang2022holistic,zhang2022opt}. \nRecently, LLMs with Reinforcement Learning from Human Feedback (RLHF) can further generate a response that is well aligned with human values~\\cite{ouyang2022training,lambert2022illustrating,bai2022training,gao2023scaling}. \nLLMs now showcase remarkable performance on a wide array of tasks, including commonsense reasoning~\\cite{zhang2022task}, text summarization~\\cite{goyal2022news, chen2022unisumm}, and massive multitask language understanding~\\cite{hendryckstest2021}. \nDifferent from previous work, our paper focuses on the task of constituency parsing, which poses a unique challenge for LLMs since its structured output is inherently divergent from plain text.", "candidate_comments": ["\\xfbai{Different from previous work, our paper focuses on the task of constituency parsing, which poses a unique challenge for LLMs since its structured output is inherently divergent from conventional text. \nWe consider both commercial models, such as ChatGPT~\\cite{BrownMRSKDNSSAA20,ouyang2022training} and GPT-4 under the in-context learning setting~\\cite{BrownMRSKDNSSAA20}, and open-source models, such as LLaMA~\\cite{touvron2023llama} under both in-context learning setting and fine-tuning settings.}"]}
{"annotations": [{"id": 86732654, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86827291, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86731843, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_9", "final_parag": "We first adapt three linearization strategies to transform constituency trees into sequences of symbols (Sec~\\ref{sec:linearization}). \nTo study the zero-shot/few-shot learning abilities of LLMs, we design prompts to guide them to generate outputs without parameter update~(\\S~\\ref{sec:prompting}). \nIn addition, we perform fine-tuning on the full training dataset to exploit the potential of open-source LLMs on constituency parsing~(\\S~\\ref{sec:tuning}).", "candidate_comments": ["Specifically, we transform them into a sequence of symbols using two linearization strategies~(Sec~\\ref{sec:linearization}). "]}
{"annotations": [{"id": 86962348, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86962359, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86970771, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_10", "final_parag": "For ChatGPT and GPT-4, we follow an instruction-prompt scheme to teach LLMs to generate linearized constituency trees~\\cite{ouyang2022training}. \nFigure~\\ref{fig:prompt} shows the detailed structure of our prompt. \nIn the zero-shot setting, our prompt contains three parts: \\texttt{Task Introduction}, \\texttt{Instruction}, and \\texttt{Task Input}. \n\\texttt{Task Introduction} explains the constituency parsing task and introduces constituent tags, ranging from clause level and phrase level to word level. \n\\texttt{Instruction} describes the task that needs to be solved and specifies the required output format. \n\\texttt{Task Input} refers to the input sentence to parse. \nIn the few-shot setting, a small number of \\texttt{Training Instances} are additionally provided as a context to instruct the LLMs to predict the constituency tree of the input sentence.", "candidate_comments": ["The input of the model consists of an instruction, an input sentence as well as the linearized tree of the input sentence."]}
{"annotations": [{"id": 86982901, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86986040, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86986024, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_11", "final_parag": "To explore the full potential of LLMs on constituency parsing, we fine-tune LLMs on the full training dataset.", "candidate_comments": ["The input of the model consists of an instruction, an input sentence as well as the linearized tree of the input sentence."]}
{"annotations": [{"id": 86864320, "completed_by": {"id": 70661}, "result": ["Différent", "Différent", "Réécriture"]}, {"id": 86901175, "completed_by": {"id": 62306}, "result": ["Réécriture", "Différent", "Différent"]}, {"id": 86901376, "completed_by": {"id": 62471}, "result": ["Différent", "Différent", "Différent"]}], "agreement": 83.33333333333334, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_12", "final_parag": "\\textbf{Post-processing} LLMs-based parsers may generate constituency trees that are not necessarily valid. \nTo this end, we add/remove parentheses to/from the generated tree to make it as valid as possible in terms of the bracket form. \nIn addition, we remove any token which is not a possible continuation given the token that precedes it. \nFurthermore, we truncate the generated tree according to the input sentence to deal with hallucinated tokens. \nIt should be noted that our post-processing approaches address content limited to a few tokens, and can not handle span-level errors.\\footnote{Our post-processing method can reduce 97.91\\% and 35.06\\% errors on in-domain and cross-domain testsets for LLaMA-7B, respectively.}", "candidate_comments": ["\\jialong{ When conducting cross-domain parsing, the low rate of valid parsing tree on ChatGPT may be attributed to the inherent invalidity of the sentences themselves.  The datasets contains numerous garbled sentences, some of which terminate with a question mark, and there are also certain sentences that involve sensitive topics, thereby causing ChatGPT to refrain from parsing them. \n}", "Appendix~\\ref{appendix:preprocess} provides a comparison of the content before and after the replacements. \nWe will provide a comparison of the content before and after the replacements in our code.", "For example, they may fail to correctly tokenize according to gold standards, exhibit missing words, or have misaligned parentheses, among other issues. \nAs the first step,  we handle parentheses by adding or removing parentheses to the generated tree to make it as valid as possible in terms of the bracket form. \nSubsequently, during evaluation, we perform post-processing on parsing tree errors. \nThis involves extracting the predicted bracket and then calculating the overall F1 score\\footnote{https://parser.yzhang.site/}. \nIt should be noted that we only address errors related to token unmatches or length unmatches. \nIf the model outputs a completely invalid constituency tree, we are unable to address it. \nMore details are provided in Appendix~\\ref{appendix:detailed}."]}
{"annotations": [{"id": 86985314, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86986630, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86989372, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_13", "final_parag": "\\textbf{Hyper-parameters}. We set the decoding temperature as 0 to increase the model’s determinism. \nFor fine-tuning, we use a learning rate of 3e-5 for 7B and 13B models, and using 1e-5 for 33B and 65B models. \nWe set the maximum length to 2048. \nDifferent from full parameter fine-tuning, we use a learning rate of 1e-4 for LoRA fine-tuning. \nWe evaluate the model performance in the validation dataset after each epoch and choose the model with the highest F1 score as the best model for testing. \nOur model is implemented based on \\textit{pytorch}\\footnote{https://github.com/pytorch/pytorch} and \\textit{huggingface transformers}\\footnote{\\url{https://github.com/huggingface/transformers}} libirary.", "candidate_comments": ["We first conduct develop experiments to study which  serialization format does LLMs prefer, and assess the impact of decoding strategies. \nWe first study which serialization format does LLMs prefer for constituency parsing. "]}
{"annotations": [{"id": 86962874, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86984102, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86984007, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_14", "final_parag": "We first investigate which serialization format is the most preferred by LLMs for constituency parsing. \nSpecifically, we take LLaMA-7B as the basic model and compare the performance of the model on the PTB test dataset.\\footnote{We have also experimented with ChatGPT, GPT-4, and LLaMA, and the overall trends are similar with LLaMA.} Figure~\\ref{fig:serilization} provides a comparison of three serialization formats in zero shot, few shot and fine tuning settings. \nIn the zero-shot setting, it can be observed that LLMs initially exhibit weak capabilities in generating all three types of trees. This is intuitive since linearized trees are different from pre-training text and the task of consistency parsing is new to LLMs.", "candidate_comments": ["We first conduct develop experiments to study which  serialization format does LLMs prefer, and assess the impact of decoding strategies. \nWe first study which serialization format does LLMs prefer for constituency parsing. "]}
{"annotations": [{"id": 86736430, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86771237, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86827005, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_15", "final_parag": "In the few-shot setting, span-based linearization gives the best result among the three variants, with an F1 score of 13.72. \nThe reason is that the span-based linearization exhibits in natural language format, which is easier for LLMs to predict than the other two methods. \nBracket-based linearization obtains the second-best result, which is about 4 points lower than span-based method.", "candidate_comments": ["\\jialong{ Due to the format of a span-based tree, which are closer to the natural language expression and  more consistent with the pre-training corpus, an F1 score of 13.92 is achieved.}"]}
{"annotations": [{"id": 86960978, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86971707, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86971955, "completed_by": {"id": 140156}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_16", "final_parag": "In contrast, the performance of LLMs is poor when they are prompted to generate transition-based trees. \nThis suggests that transition-based trees are harder to generate than bracket-based linearization and span-based linearization.", "candidate_comments": ["\\jialong{ Due to the format of a span-based tree, which are closer to the natural language expression and  more consistent with the pre-training corpus, an F1 score of 13.92 is achieved.}"]}
{"annotations": [{"id": 86904510, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86975136, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86975269, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_17", "final_parag": "We then study the impact of different decoding strategies during inference.  Figure~\\ref{fig:search} compares the performance of greedy search and beam search with beam sizes ranging from 2 to 5, on the development dataset of the PTB. In general, the decoding strategy does not impact the parsing performance. \nIn particular, for LLaMA-7B, greedy search yields an F1 score of only 0.1 lower than beam search with a beam size of 5. This suggests that the decoding strategy is not the bottleneck for LLM-based constituency parsing. \nMoreover, the impact of the decoding strategy on LLaMA-33B has a similar trend compared to LLaMA-7B.\\footnote{We do not report the results of LLaMA-65B since it exceeds GPU memory limit when beam size $\\geq$ 4.}", "candidate_comments": ["Compared with state-of-the-art chart-based and transition-based systems, LLaMA-65B gives relatively lower results. \nThis indicates that generatively-trained LLMs can not solve constituency parsing."]}
{"annotations": [{"id": 86725636, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86728116, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86730136, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_18", "final_parag": "\\subsection{The Capacity of LLMs on Constituency Parsing} To study the full potential of LLMs on constituency parsing, we fine-tune LLMs on the full training dataset of PTB and compare the performance of LLMs with the state-of-the-art method. \nTable~\\ref{table:finetune} presents the results of different systems on the PTB test set. Among all LLM-based methods, LLaMA-65B achieves the highest result, with an F1 score of 95.90. \nCompared with sequence-based baselines, fine-tuned LLMs yield greatly superior results. Notably, LLaMA-65B gives an improvement of approximately 2.2 F1 score over GPT-2. This observation suggests that LLMs can substantially boost the performance of sequence-based constituency parsers.", "candidate_comments": ["Compared with state-of-the-art chart-based and transition-based systems, LLaMA-65B gives relatively lower results. \nThis indicates that generatively-trained LLMs can not solve constituency parsing."]}
{"annotations": [{"id": 86963025, "completed_by": {"id": 62471}, "result": ["Réécriture", "Différent"]}, {"id": 86984696, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86984813, "completed_by": {"id": 70661}, "result": ["Réécriture", "Différent"]}], "agreement": 77.77777777777777, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_19", "final_parag": "Compared with the state-of-the-art chart-based and transition-based systems, LLaMA-65B gives competitive results. Specifically, LLaMA-65B outperforms the SEPar parser's 95.72 F1 score, narrowing the gap between the sequence-based parser and chat-based parser. \nThis indicates that the LLM-based parser can serve as a strong backbone for constituency parsing. \nLLaMA-65B produces relatively lower results compared with SAPar and TGCN, the reason is that SAPar and TGCN employ external mechanisms to learn local features.", "candidate_comments": ["Compared with state-of-the-art chart-based and transition-based systems, LLaMA-65B gives relatively lower results. \nThis indicates that generatively-trained LLMs can not solve constituency parsing.", "Regarding to model scale, LLaMA-13B outperforms LLaMA-7B by about 0.18 F1 score while LLaMA-33B outperforms LLaMA-13B by 0.32 F1 score, and LLaMA-65B gives 0.1 F1 improvements over LLaMA-33B. \nThis indicates that larger scale can bring better performance."]}
{"annotations": [{"id": 86984367, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985412, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86985544, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_20", "final_parag": "Regarding the model scale, LLaMA-13B improves LLaMA-7B by approximately 0.18 in F1 score, while LLaMA-33B surpasses LLaMA-13B by 0.32 in F1 score. \nMoreover, LLaMA-65B further delivers a 0.09 F1 score improvement over LLaMA-33B. \nThese results suggest that increasing the model scale can still bring performance improvements even with billions of parameters.", "candidate_comments": ["Regarding to model scale, LLaMA-13B outperforms LLaMA-7B by about 0.18 F1 score while LLaMA-33B outperforms LLaMA-13B by 0.32 F1 score, and LLaMA-65B gives 0.1 F1 improvements over LLaMA-33B. \nThis indicates that larger scale can bring better performance."]}
{"annotations": [{"id": 86973348, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture", "Réécriture"]}, {"id": 86981669, "completed_by": {"id": 70661}, "result": ["Différent", "Différent", "Différent"]}, {"id": 86965916, "completed_by": {"id": 140156}, "result": ["Réécriture", "Différent", "Différent"]}], "agreement": 50.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_21", "final_parag": "As shown in the last two rows of Table~\\ref{table:finetune}, both instruction-tuned models deliver slightly lower performance than vanilla LLMs. This indicates that instruction-tuning cannot benefit constituency parsing under the fine-tuning setting. \nThe reason can be that instruction-tuning tasks primarily focus on semantics, resulting in forgetting of constituent knowledge.", "candidate_comments": ["Compared with standard fine-tuning, LoRA-tuning gives relatively lower results. \nFor example, LLaMA-7B [\\texttt{L}] gives 2.3 lower F1 score compared with LLaMA-7B.  This is contradictory to the existing observations~\\cite{} that LoRA-tuning can give slightly lower or comparable results than fine-tuning.  The reason can be that the output format of constituency parsing is more complex than other text-based tasks, thus requiring more parameter changes.", "\\jialong{This indicates that the amount of instruction data is insufficient, with Alpaca utilizing only 52k and Vicuna using 70k. \nFurthermore, the form of instruction does not align well with our task, resulting in limited acquisition of consistent knowledge.}", "Compared with standard fine-tuning, LoRA-tuning gives relatively lower results. \nFor example, LLaMA-7B [\\texttt{L}] gives 2.3 lower F1 score compared with LLaMA-7B.  This is contradictory to the existing observations~\\cite{} that LoRA-tuning can give slightly lower or comparable results than fine-tuning.  The reason can be that the output format of constituency parsing is more complex than other text-based tasks, thus requiring more parameter changes."]}
{"annotations": [{"id": 86736815, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86959353, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86960427, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_22", "final_parag": "This observation is different from existing findings on sentence classification and text generation tasks~\\cite{HuSWALWWC22,xu2023baize} which suggest LoRA-tuning can deliver results that are either slightly lower or comparable to those achieved through fine-tuning. A potential reason could be that the output format of constituency parsing is more complex than that of other text-based tasks, thereby necessitating more substantial parameter modifications.", "candidate_comments": ["Compared with standard fine-tuning, LoRA-tuning gives relatively lower results. \nFor example, LLaMA-7B [\\texttt{L}] gives 2.3 lower F1 score compared with LLaMA-7B.  This is contradictory to the existing observations~\\cite{} that LoRA-tuning can give slightly lower or comparable results than fine-tuning.  The reason can be that the output format of constituency parsing is more complex than other text-based tasks, thus requiring more parameter changes."]}
{"annotations": [{"id": 86972531, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86981574, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86981606, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_23", "final_parag": "Table~\\ref{table:fewshot} documents the zero-shot and 5-shot learning performance of different LLMs on the testset of PTB.\\footnote{Due to the difficulty of conducting large-scale evaluations with GPT-4, we conduct evaluations on a subset of 100 samples which is randomly selected from original testsets.} It is noteworthy that commercial LLMs (ChatGPT and GPT-4) deliver impressive results in both scenarios, achieving approximately 60 and 70 in F1 scores in zero-shot and 5-shot settings, respectively. \nIn contrast, open-source LLMs show much worse performance. Specifically, all open-source LLMs deliver low results in the zero shot setting and give less than 30 F1 scores in the 5-shot setting. \nThis suggests that the syntactic knowledge acquired during pre-training cannot be directly transferred to the constituency parsing task, thereby confirming our assumption. \nAdditionally, LLMs show better performance as model parameters increase, suggesting a positive correlation between generalization ability and model size.", "candidate_comments": ["We study the generalization ability in both in-domain and cross-domain settings. \nWe first evaluate the task generation ablity of LLMs on the test set of PTB.  Table~\\ref{table:fewshot} reports the zero-shot and five-shot learning performance of different LLMs. \nNotably, commercial LLMs (ChatGPT and GPT-4) gives impressive results in both settings, with about 60 and 70 F1 score in zero-shot and five-shot scenario. \nIn contrast, public LLMs report much worse performance.  In particular, all public LLMs fail in zero-shot setting and reports less than xx F1 score in five-shot setting. \nThis suggest that the syntactic knowledge learned in pre-training can not be directly transfered to constituency parsing task, which confirms our assumption."]}
{"annotations": [{"id": 86960928, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973199, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974563, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_24", "final_parag": "To study the cross-domain generalization capability of LLMs, we evaluate the zero-shot domain adaptation performance in five cross-domain test sets. \nTable~\\ref{table:cross-domain} shows the F1 score on different datasets and the relative performance reduction rate. Remarkably, ChatGPT and GPT-4 obtain impressive performance without training. \nCompared with the ChatGPT series, the fine-tuned models (SEPar, SAPar, GPT-2, LLaMA) give relatively better F1 scores across all cross-domain sets.", "candidate_comments": ["To study the generalization of LLMs across different domains, we conduct evaluations on five cross-domain datasets. \nTable~\\ref{table:cross-domain} lists the performance of LLMs on the cross-domain datasets. \nRegardless of few-shot setting or a full tuning setting, the overall performance of the model on cross-domain tasks exhibits a significant decrease compared to in-domain task. In particular, xxx ChatGPT series xxx. \nCompared with ChatGPT series, the fine-tuned LLaMAs give relatively better results in all testsets. \nCompared with GPT, LLaMAs achieve lower relative performance reduction rate on cross-domain testsets, indicating that LLMs can improve the cross-domain generalization ability of constituency parsers. \nMoreover, the relative performance reduction rate of LLaMAs are larger compared with SEPar parser, the reason can be that xxxx."]}
{"annotations": [{"id": 86960481, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}, {"id": 86974998, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86972707, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}], "agreement": 55.55555555555556, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_25", "final_parag": "Compared with GPT-2, LLaMAs show relatively lower performance reduction rates, and these reduction rates decrease as the model parameters increase. \nThis suggests that larger models have relatively better generalization ability for domain shifting.", "candidate_comments": ["This suggests that all tested models still have room to improve their cross-domain generalization ability.", "This suggests that LLMs \\textbf{does not} have significantly better cross-domain generalization ability than medium-sized models, even seeing more training data."]}
{"annotations": [{"id": 86983885, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86986543, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86986715, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_26", "final_parag": "Different from chart-based models, a drawback of sequence-based models, including LLMs, is that there is no mechanism, such as the CKY algorithm~\\cite{pereira1983parsing} or using a state buffer~\\cite{yamada-matsumoto-2003-statistical}, which can ensure the validity of the generated constituency parse trees. \nTo study the impact of invalid generated trees, we calculate the valid F1 and invalid rate of fine-tuned LLaMA-7B on the testset of PTB and compare them with SEPar. \nAs shown in Table~\\ref{appendix:hurt}, invalid trees greatly hurt the overall performance of LLaMA-7B. In particular, the F1 score on the law domain is 90.69 for valid trees, which is about 4 points higher than the overall performance. \nIn contrast, the invalid rates of SEPar are consistently zero in all testsets. \nIn addition, the invalid rates of cross-domain testsets are higher compared to in-domain testset, the reason is that cross-domain testsets contain novel patterns which do not exist in the in-domain training dataset.", "candidate_comments": ["Furthermore, we observe that the performance of the ChatGPT series on cross-domain datasets is not as strong as their performance on the PTB dataset. \nThe cross-domain datasets contains numerous garbled sentences, some of which terminate with a question mark, and there are also certain sentences that involve sensitive topics, thereby causing ChatGPT series to refrain from parsing them. \n\\subsection{Can Instruction Tuning helps Constituency Parsing?}"]}
{"annotations": [{"id": 86962296, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86964712, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86965006, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_27", "final_parag": "To further understand the problem of invalid tree generation, we divide invalid trees into different groups and analyze the distribution of each group, and the results are presented in Table~\\ref{table:invalid}.", "candidate_comments": ["In the \\textit{EVALB} evaluation script, invalid trees are categorized into three types: word mismatch, length mismatch, and prediction failure."]}
{"annotations": [{"id": 86982911, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86985601, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86985661, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_28", "final_parag": "Following the \\textit{EVALB} evaluation script, we categorize invalid trees into three types: length mismatch, word mismatch, and prediction failure.", "candidate_comments": ["In the \\textit{EVALB} evaluation script, invalid trees are categorized into three types: word mismatch, length mismatch, and prediction failure."]}
{"annotations": [{"id": 86962798, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86977554, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86979005, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_29", "final_parag": "We further categorize length mismatch errors into four subcategories: repetition, task-dofocus, and other errors. \nAmong them, ``repetition'' is frequently observed in generative models, which refers to the phenomenon where model outputs contain large repeated contents~\\cite{holtzman2019curious}. \n``Task-defocus'' refers to the phenomenon that LLMs modify the given sentence before performing constituency parsing. \n``Others'' encompasses errors such as incorrect word segmentation or missing words. \nIn addition, word mismatch accounted for a quarter of the cause of invalid trees. The number of word mismatch errors is larger in cross-domain testsets compared with in-domain testsets. \nFinally, prediction failures account for a small proportion of invalid trees. \nOur human analysis shows that prediction failures are typically caused by informal input sentences or phases.", "candidate_comments": ["We can observe that LLMs still make numerous errors in constituency parsing task, stemming from limited ability to precisely control the output of generative models, even after fine-tuning on a dataset of 40k examples. \nThe main reason can be the informal use of language, including punctuations, non-English characters, typos, and less structured sentences or phrases."]}
{"annotations": [{"id": 86862416, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86900853, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86900954, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_30", "final_parag": "Figure~\\ref{fig:length} shows the performance of tree models against input length. \nAll three models give relatively lower results when parsing long input sentences. \nGPT-2 shows the largest performance reduction among the three models. \nCompared with GPT-2, LLaMA-7B sees lower performance decreases on long sentences, showing that LLMs are more robust to long sentences. Compared with SEPar, LLaMA-7B gives comparable performance decreases when input sentences have less than 40 tokens, while the gap is larger when parsing longer sentences. \nThis indicates that LLMs are \\textbf{unable} to fundamentally address the performance limitation of sequence-based parsers.", "candidate_comments": ["Figure~\\ref{fig:length} shows the reduction between model performance evaluated on in-domain and out-of-domain data. \nSimilarly, we evaluate SEPar parser and LLaMA-7B (tuned on parsing data) on sentences with token lengths $\\geq$40. \nThe performance drop between F1 score for sentences $\\geq$ 40 and overall F1 score in performance of LLaMA is much severer compared to the SEPar Parser. \nThis indicates that the generative LLaMA is not as effective as the bidirectional SEPar Parser in the context of long-text constituency parsing.  This observation highlights one of the limitations of generative models."]}
{"annotations": [{"id": 86721567, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86730841, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86765703, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_31", "final_parag": "The results are shown in Figure~\\ref{fig:data_effiency}, where we report the overall F1 score and F1 score of valid trees for LLaMA-7B and GPT-2. \nSince chart-based system SAPar does not generate invalid trees, resulting in the overall F1 score and the F1 score of valid trees being the same.", "candidate_comments": ["The results of this exploration are presented in The comparative analysis between LLaMA-7B and GPT-2 was conducted to assess data efficiency, and the results are shown in Figure~\\ref{fig:data_effiency}."]}
{"annotations": [{"id": 86864034, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86902172, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86903302, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19462-2c90f92f6f84ce60.jsonl", "id_parag": "2310.19462-2c90f92f6f84ce60_32", "final_parag": "Different from previous observations~\\cite{chen2023maybe}, LLaMA-7B does not reach its peak performance until using full training data.\\footnote{We have also tried to train models with silver data, without observing great improvements.} The reason can be that the structured output of constituency parsing is more challenging to generative LLMs compared with other text-based tasks. \nIn addition, LLaMA-7B outperforms GPT-2 in all different training settings, especially when there are less than 5000 training instances. \nAlso, LLaMA-7B gives better results than SAPar when the number of training instances is less than 1000. \nThis shows that LLMs have better generalization abilities compared to medium-size models. \nWhen the size of training instances increases, the performance gap between LLaMA-7B and GPT-2 becomes less. \nFurthermore, both models suffer greatly from invalid trees when trained with limited data, and the gap between overall F1 and valid F1 becomes less when the number of training instances increases. \nCompared with GPT-2, LLaMA-7B sees a smaller performance gap, indicating that LLMs have better abilities to generate a valid tree.", "candidate_comments": ["Similarly, the gap between overall model performance and model performance of valid trees becomes less when the number of training instances increases. \nNotably, LLaMA-7B's upward trend outperforms that of GPT-2."]}
{"annotations": [{"id": 86904589, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86985530, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86985584, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19470-507e18d9ae51c8ef.jsonl", "id_parag": "2310.19470-507e18d9ae51c8ef_1", "final_parag": "Additionally, to verify that ``Grokking ticket'' are a more critical factor than weight norms, we compared the ``good'' subnetworks with a dense network having the same L1 and L2 norms. Results show that the subnetworks generalize faster than the controlled dense model. \nIn further investigations, we discovered that at an appropriate pruning rate, grokking can be achieved even without weight decay.", "candidate_comments": ["We also show that the speedup is significant even when compared with the dense networks with the same weight norm as the lottery tickets.  In addition, to verify that ``Grokking tickets'' are a more crucial factor than weight norms, we compared the ``good'' subnetworks with the same L1 and L2 norm dense network, and results show that the subnetworks generalize faster than the controlled dense model."]}
{"annotations": [{"id": 86962741, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973095, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974359, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19470-507e18d9ae51c8ef.jsonl", "id_parag": "2310.19470-507e18d9ae51c8ef_2", "final_parag": "We also show that speedup does not happen when using tickets identified at the memorization solution or transition between memorization and generalization or when pruning networks at the initialization (Random pruning, Grasp, SNIP, and Synflow). The results indicate that the weight norm of network parameters is not enough to explain the process of grokking, but the importance of finding good subnetworks to describe the transition from memorization to generalization. The implementation code can be accessed via this link: \\url{https://github.com/gouki510/Grokking-Tickets}.", "candidate_comments": ["We also show that the speedup is significant even when compared with the dense networks with the same weight norm as the lottery tickets.  In addition, to verify that ``Grokking tickets'' are a more crucial factor than weight norms, we compared the ``good'' subnetworks with the same L1 and L2 norm dense network, and results show that the subnetworks generalize faster than the controlled dense model."]}
{"annotations": [{"id": 86985967, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 86986569, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86990186, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19470-507e18d9ae51c8ef.jsonl", "id_parag": "2310.19470-507e18d9ae51c8ef_3", "final_parag": "In the experiments on the modular addition, we used two architectures: (1) MLP and (2) Transformer. In both architectures, we encode the $a$ and $b$ into one hot vector. For both architectures, the cross entropy loss is used.", "candidate_comments": ["In this paper, we investigate grokking using two type of tasks (Modular addition and MNIST) and two architectures (MLP and Transformer), to show the universality of the grokking tickets.  There are studies that analyze grokking from the structure of the Transformer (\\cite{nanda2023progress}) and studies that conduct specific analyses for specific tasks (\\cite{merrill2023tale}). However, our research does not depend on the task or architecture because The concept of lottery tickets has been verified across various fields.  \\"]}
{"annotations": [{"id": 86971169, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86982026, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86982598, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19470-507e18d9ae51c8ef.jsonl", "id_parag": "2310.19470-507e18d9ae51c8ef_4", "final_parag": "\\subsubsection{MNIST classification} Following \\citep{liu2023omnigrok,varma2023explaining} we split the dataset into 1k training examples and 60k testing examples.", "candidate_comments": ["It is known that grokking also occurs in MNIST class classification by \\cite{liu2023omnigrok}, \\cite{varma2023explaining}.  We follow the experimental setup used in the prior studies.  Specifically, we split training and test data in to 1k:60k. "]}
{"annotations": [{"id": 86986118, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86987372, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86989422, "completed_by": {"id": 70661}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19470-507e18d9ae51c8ef.jsonl", "id_parag": "2310.19470-507e18d9ae51c8ef_5", "final_parag": "In \\autoref{exp:gro vs base}, we only tested when the pruning rate was $0.6$. \nThe left panel of \\autoref{fig:effect_of_k} depicts the generalization speed across various pruning rates, while the right panel demonstrates the correlation between pruning rate and grokking speed. \nWe use the grokking tickets in this setup. The task is modular addition and MLP architecture. A pruning rate of $0.0$ represents the base model because no weights are pruned and the model is rewound to its initial weights.", "candidate_comments": ["In this section, we will verify how the grokking speed changes when the pruning rate is altered. In Fig \\ref{pr_test_acc}, pruning rate 0 line is the original grokking phenomenon.It is observed that when LT is discovered with a pruning rate ranging from 0.1 to 0.7, grokking is accelerated. However, when the pruning rate is 0.9, meaning only 10\\% of the weights are being used, the grokking phenomenon is not observed. It is presumed that this is because the pruning has extended even to the weights included in the grokking tickets. "]}
{"annotations": [{"id": 86962781, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86973165, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974340, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19470-507e18d9ae51c8ef.jsonl", "id_parag": "2310.19470-507e18d9ae51c8ef_6", "final_parag": "We can make the following observations: (1) Most pruning rato (0.1, 0.3, and 0.7) accelerate the grokking, indicating that the above observation does not heavily depend on the selection of the pruning rato. When the pruning rate is 0.7, the speed of grokking is the fastest, according to the right figure. (2) At the same time, with a higher pruning rate, such as 0.9, the grokking phenomenon is not observed, which might be due to the lack of capacity.", "candidate_comments": ["In this section, we will verify how the grokking speed changes when the pruning rate is altered. In Fig \\ref{pr_test_acc}, pruning rate 0 line is the original grokking phenomenon.It is observed that when LT is discovered with a pruning rate ranging from 0.1 to 0.7, grokking is accelerated. However, when the pruning rate is 0.9, meaning only 10\\% of the weights are being used, the grokking phenomenon is not observed. It is presumed that this is because the pruning has extended even to the weights included in the grokking tickets. "]}
{"annotations": [{"id": 86972316, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86970675, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86972284, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19470-507e18d9ae51c8ef.jsonl", "id_parag": "2310.19470-507e18d9ae51c8ef_7", "final_parag": "In this case (\\autoref{fig:grokking_without_wd}-a), we recognize that the critical pruning rate is between 0.8 and 0.9 because if the pruning rate increases to 0.9, the test accuracy dramatically decreases. \nThus, we gradually increased the pruning rate in increments of 0.01 from 0.8 and found that the $k=0.81$ is the critical pruning ratio. We then compare the behavior of the grokking tickets with the critical pruning rato with ($\\alpha=1.0$) and without weight decay ($\\alpha=0.0$). \n\\autoref{fig:grokking_without_wd}(b) show the results of the experiments. For the baseline, we also plot the test accuracy when using the base model. As shown in the figure, the test accuracy reaches perfect generalization almost similarly with and without weight decay. The results show that the grokking tickets with the critical pruning ratio do not require any weight decay during the optimization, indicating that the weight decay pushes the networks to find good lottery tickets during the transition between memorization and generalization, rather than just decreasing the weight norm of the networks. For experiments on the transformer, refer to the \\autoref{exp:trans_exp5}.", "candidate_comments": ["As mentioned in \\ref{sec:pruning rate}, \\autoref{fig:grokking_without_wd}(a)) show that if the pruning rate is raised too much (0.9), the test accuracy does not improve at all because even the good weight is pruned. On the other hand, if the pruning rate is low (0.1, 0.2, 0.3), the test accuracy also does not improve because many redundant edges remain in the model since there is no weight decay. ", "As mentioned in \\ref{sec:pruning rate}, it has been confirmed that if the pruning rate is raised too much, the test accuracy does not improve at all. To find the maximum pruning rate without lowering the test accuracy。 we observed the test accuracy when changing the pruning rate by 0.01. It can be seen that around p=0.81 is the maximum where test accuracy does not decrease. Hence, we define this as the critical prune rate.\\\\ In Fig \\ref{without decay}, It can be observed that grokking occurs even in the absence of decay.\\\\ This result implicitly suggests that the weight decay used in grokking has been employed for the discovery of good subnetwork. \nHowever, in Fig \\ref{without decay},  The answer is that the subnetworks found at larger pruning rates are more redundant compared to subnetwork found at the critical pruning rate. In other words, they still have weights that should be pruned."]}
{"annotations": [{"id": 86904283, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86972510, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86972742, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19470-507e18d9ae51c8ef.jsonl", "id_parag": "2310.19470-507e18d9ae51c8ef_8", "final_parag": "For the second question Q2, we compared the grokking tickets with the subnetworks identified differently. Specifically, we tested three well-known pruning at initialization methods (Grasp~\\citep{wang2020picking}, SNIP~\\citep{lee2019snip}, and Synflow~\\citep{tanaka2020pruning}) and random pruning as baseline methods. The basic flow of the experiment is as follows: \\begin{enumerate} \\item Randomly initialize a neural network $f(\\vx;\\vtheta_0)$. \\item Prune \\textit{p}\\% of the parameters in $\\vtheta_{0}$ according to the scores $S(\\vtheta)$, creating a mask m . \\item Train the network from $\\vtheta_0 \\odot \\vm$. \\end{enumerate} According to \\citet{tanaka2020pruning}, research on pruning at initialization boils down to the methodology of determining the score in the above process 2. \nFor details on each of the pruning methods, refer to the section \\ref{detail prunig at init}.", "candidate_comments": ["$ %\\gm{TODO; other pruning method} $ %Next, we investigated the effects of pruning methods on Grokking. \n$ %\\subsubsection{Methods other than magnitude pruning} $ %\\gm{IMP} $ In the aforementioned sections, subnetwork were determined using trained weights. This method requires intentionally over-parameterizing  model and training it, offering limited practical advantages. If generalized lottery tickets (like grokking tickets ) could be acquired before training, it would be possible to converge to generalized solutions without going through memorized ones, proving beneficial from the perspective of learning efficiency. \n$ Therefore, we have attempted existing methods that prune weights at initialization. The flow of the experiment is as follows: $ \\begin{enumerate} $     \\item Randomly initialize a neural network $f(\\vx;\\vtheta_0)$.   $     \\item Prune \\textit{p}\\% of the parameters in $\\vtheta_{0}$ according to the scores $S(\\vtheta)$, creating a mask $\\vm$ .  $     \\item Train the network from $\\vtheta_0 \\odot \\vm$.   $ \\end{enumerate} $ According to \\cite{tanaka2020pruning}, research on pruning at initialization boils down to the methodology of determining the score in the above process 2.  $ We experimented with four different pruning methods, Synflow (\\cite{tanaka2020pruning}), Grasp(\\cite{wang2020picking}), SNIP(\\cite{lee2019snip}).  $  For comparison purposes, we also examined Magnitude Pruning at initialization, random pruning, grokking tickets and Base model(dense network). In each pruning method, we use same initial weights.\\\\ In Fig \\ref{pruning at init}, the methods of pruning at initialization all resulted in grokking speeds that were not only slower than the Base Model but also yielded results that were comparable to, if not worse than, those of models pruned randomly. \\\\ Consistent with the results of the section \\ref{exp;rewind}, this result implies that merely pruning does not accelerate grokking, meaning it does not promote finding the generalized solutions more swiftly. "]}
{"annotations": [{"id": 86962091, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973058, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86973598, "completed_by": {"id": 126844}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19470-507e18d9ae51c8ef.jsonl", "id_parag": "2310.19470-507e18d9ae51c8ef_9", "final_parag": "\\subsection{Sparsity and Efficiency} \\citet{merrill2023tale} argued that the grokking phase transition corresponds to the emergence of a sparse subnetwork that dominates model predictions. \nThey empirically studied the internal structure of networks undergoing grokking on sparse parity tasks. The objectives of this study are similar to those of our research. \\citet{merrill2023tale} has only conducted experiments in n-k parity and has further verified it in tasks where sparsity is inherently apparent. On the other hand, we are conducting a more generalized analysis of grokking, independent of the task or architecture.", "candidate_comments": ["From the perspective of weight norms, grokking seems to be induced by weight decay; however, when refining the weight decay as in our experiment, it was found that there is a causal relationship between grokking and the discovery of Lottery Tickets."]}
{"annotations": [{"id": 87255480, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87259084, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87268327, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19470-507e18d9ae51c8ef.jsonl", "id_parag": "2310.19470-507e18d9ae51c8ef_10", "final_parag": "\\citet{liu2022understanding} analyzed that the dynamics of representation learning and the dependency on the training set size have been predicted, demonstrating that the origin of grokking lies in learning structured representations. \n\\citet{nanda2023progress} comprehends the underlying training dynamics in grokking and can divide the training into three phases: memorization of training data, circuit formation, and cleanup. \nThe relationship between representation learning and the discovery of lottery tickets is still not well understood. In \\autoref{apen:grokking representation}, following to \\citet{google_blogpost}, we investigated the periodicity of each neuron by plotting input-side weight and output-side weight of each neuron. \nThe results demonstrate that grokking tickets contribute to the acquisition of a good representation.", "candidate_comments": ["\\gm{If there is research about lottery tickets and model size, it may be related to the relationship between grokking and double descent or emergent ability} % From the perspective of lottery tickets, there are a lot of noise weight in the network."]}
{"annotations": [{"id": 86962223, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86962280, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86964570, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19503-66c30bd111f2a070.jsonl", "id_parag": "2310.19503-66c30bd111f2a070_1", "final_parag": "\\added{\\textbf{Grounding based on our Illustrative Scenario.}} \\added{Figure \\ref{fig:MR-policies} illustrates how machine-readable policies and norms can be used to support self-determination. \nConsidering our illustrated scenario} individuals may want to establish policies to precisely define the subset of their PKGs to be shared with communities and what forwarding they allow. For example, \\emph{share with the diabetes community my blood in sugar values measured by my connected device and the output of my AI healthcare assistant, or only share and forward anonymised aggregates to medical research institutions, or contact me for negotiation if the pharma company is interested in using my data for clinical studies.} Communities may do the same, \\emph{e.g}, requiring \\added{specific data to be shared} to join the community, but also \\added{requiring agreements in order to ensure that participants will abide by} social and behavioural norms needed for self-regulation. Public and private organisations may need to adhere not only to privacy preferences and licenses but also to various \\added{general regulations, e.g., the GDPR, the proposed AI Act in the EU or the Health Insurance Portability and Accountability Act (HIPAA)\\footnote{\\url{https://www.hhs.gov/hipaa/index.html}} in the US, as well as domain-specific regulations (e.g., advanced therapy medicinal products\\footnote{\\url{https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex\\%3A32007R1394}} and rare diseases\\footnote{\\url{https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex\\%3A32009H0703\\%2802\\%29}}).}", "candidate_comments": ["Merged with the beginning of the section to ground in scenario and abide to agreed section structure.  Considering our running example whereby PKGs contain sensitive personal data, individuals need to be able to specify privacy preferences in terms of which data may be used, by whom, and for what purpose(s). From a community perspective, individuals need to be able to determine which data they wish to share with other members of the community, while social norms are needed in order to ensure that communities are able to self regulate not only data and service access but also behaviour. In turn, doctors may wish to share knowledge, while at the same time protecting their intellectual property. Whereas, service providers in the form of public and private organisations may need to adhere not only to privacy preferences and licenses but also to various general, e.g., the GDPR, the proposed AI Act in the EU and Health Insurance Portability and Accountability Act (HIPAA)~\\cite{act1996health} in the US, and domain specific regulations (e.g., advanced therapy medicinal products \\footnote{https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex\\%3A32007R1394}, action in the field of rare diseases\\footnote{https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex\\%3A32009H0703\\%2802\\%29}). In each case there is a need for the encoding of policies in a machine understandable format such that bots and assistants can ensure enforcement, compliance checking, and transparency. "]}
{"annotations": [{"id": 86962037, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86982854, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86983057, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19503-66c30bd111f2a070.jsonl", "id_parag": "2310.19503-66c30bd111f2a070_2", "final_parag": "This paper presents a compelling argument for integrating KG-based AI to empower individuals' self-determination and benefit society. This overarching goal is supported by three fundamental pillars: trust, accountability, and autonomy. We advocate the foundations of these pillars require focused research in four areas: machine-readable norms and policies, decentralised infrastructure, decentralised KG management, and explainable neuro-symbolic AI. By drawing on a concrete scenario within the healthcare domain, we demonstrate the relevance of each foundational topic and outline a comprehensive research agenda for each of them.", "candidate_comments": ["In this paper, we motivate the need for KG-Based AI to be a force for individuals' self-determination and the benefit of society. We posit this goal is built upon three pillars: trust, accountability and autonomy, that in turn are achieved by researching on four foundational topics: Machine Readable norms and policies, Decentralised Infrastructure, Decentralised KG Management and Explainable Neuro-Symbolic AI. Grounding on an illustrative scenario in the healthcare domain, we describe how each foundational topic related to the pillars and propose a research agenda for each of them.  "]}
{"annotations": [{"id": 86984669, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86985275, "completed_by": {"id": 140156}, "result": ["Différent"]}, {"id": 86985285, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19509-f8804a7cd3a7888b.jsonl", "id_parag": "2310.19509-f8804a7cd3a7888b_1", "final_parag": "ByteDance Inc \\\\ {\\tt\\small \\{liusongwei.zju, xuhaitao.henry, xuyuyang.nicole, wangshuai.hust, lijiashi\\}@bytedance.com}\\\\ {\\tt\\small \\{yanchenqian.i, lilianqiang, fulean, panxin.321\\}@bytedance.com}\\\\", "candidate_comments": ["{\\small yanchenqian.i, lilianqiang, fulean, panxin.321\\}@bytedance.com}\\\\ {\\tt\\small \\{baiyang.87, chengshu.2019, fulean\\}@bytedance.com} }"]}
{"annotations": [{"id": 86962450, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86977195, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86978866, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19512-223cbaab0a8fdaa3.jsonl", "id_parag": "2310.19512-223cbaab0a8fdaa3_1", "final_parag": "Currently, several open-source T2V models exist, \\textit{i.e.,} ModelScope~\\cite{wang2023modelscope}, Hotshot-XL~\\cite{hotshotxl}, AnimateDiff~\\cite{guo2023animatediff}, and Zeroscope V2 XL~\\cite{zeroscope}. \nThe released ModelScope model can only generate videos with a resolution of $256 \\times 256$, and the image quality is unsatisfactory. \nZeroscope V2 XL improves its visual quality by tuning it on a small set of videos, but flickers and visible noise persist in its generated videos. \nHotshot-XL aims to extend SDXL into a video model and produce a gif with 8 frames and a resolution of $512\\times 512$. \nAnimateDiff proposes to combine the temporal module with the spatial module of a LORA SD model. Since the temporal module is trained on Webvid-10M, the results of the original T2V model of AnimateDiff are poor. The combination with a high-quality LORA model can generate high-quality videos.", "candidate_comments": ["Although the results of the original T2V model are poor, the combination can yield high-quality videos. "]}
{"annotations": [{"id": 86972342, "completed_by": {"id": 62306}, "result": ["Différent"]}, {"id": 86982382, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86982417, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19512-223cbaab0a8fdaa3.jsonl", "id_parag": "2310.19512-223cbaab0a8fdaa3_2", "final_parag": "However, the scope is restricted by the LORA model in terms of style and concept composition ability. \nThere is still a lack of an open-source generic T2V foundation model capable of generating high-resolution and high-quality videos.", "candidate_comments": ["Although the results of the original T2V model are poor, the combination can yield high-quality videos. "]}
{"annotations": [{"id": 86971682, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86981694, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86980810, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19513-66c30bd111f2a070.jsonl", "id_parag": "2310.19513-66c30bd111f2a070_1", "final_parag": "To filter the OAS antibodies, we first remove any duplicate entries with identical concatenated CDR sequences. For filtering the SAbDab antibody complexes, we remove antibodies that have both an identical concatenated CDR sequence and epitope sequences with greater than 90\\% similarity by CD-HIT~\\cite{cdhit}. Here the epitope is defined as the residues in each antigen chain with backbone atoms within 6~Å of the antibody backbone.", "candidate_comments": ["To filter the OAS antibodies, we first remove any duplicate entries with identical concatenated CDR sequences. For filtering the SAbDab antibody complexes, we remove duplicates where the concatenated CDR sequences are identical, and the epitope sequences are clustered together with 90\\% similarity by CD-HIT~\\cite{cdhit}. Here the epitope is defined as the residues in each antigen chain with backbone atoms within 6~Å of the antibody backbone. \nTo filter the OAS antibodies, we first remove any duplicate entries with identical concatenated CDR sequences. For filtering the SAbDab antibody complexes, we remove duplicates where the concatenated CDR sequences are identical, and the epitope sequences (defined by the residues in each antigen chain with backbone atoms within 6~Å of the antibody backbone) are clustered together with 90\\% similarity by CD-HIT~\\cite{cdhit}."]}
{"annotations": [{"id": 86863092, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86901610, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86902233, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19513-66c30bd111f2a070.jsonl", "id_parag": "2310.19513-66c30bd111f2a070_2", "final_parag": "We fine-tune our AbMPNN model starting from the original ProteinMPNN model weights, first on the OAS dataset, then on the SAbDab dataset. Both of these fine-tuning steps use an Adam optimiser. We reduce the learning rate by a factor of 10 if the validation loss does not improve for 10 epochs (5 for the OAS data), starting from an initial learning rate of $5 \\times 10^{-4}$ for the OAS fine-tuning and of $10^{-4}$ for the SAbDab training step. Each epoch consists of 1000 randomly selected antigen binding fragments.", "candidate_comments": ["For training, we limit ourselves to fine-tuning from the original ProteinMPNN models with backbone noise of 0.02~Å~\\cite{proteinmpnn}. We fine-tune three different models. One model from the OAS dataset, one model from the SAbDab dataset, and a third model fine-tuned from the OAS model on the SAbDab dataset. For each training run use an Adam optimiser with the pytorch ReduceLROnPlateau scheduler that tracks the validation loss and reduces the learning rate by a factor 10 on plateau. We trial four different learning rates $5\\times 10^{-4},~ 1\\times10^{-4},~ 5\\times10^{-5}, 1\\times 10^{-5}$ and select the best model based on the minimum validation loss across all runs."]}
{"annotations": [{"id": 86865001, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86900976, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86901275, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_1", "final_parag": "Nowcasting is crucial for weather-dependent decision-making, ranging from everyday occurrences with high socio-economic implications like aviation and outdoor activities, to extreme events requiring disaster management.", "candidate_comments": ["Nowcasting is an integral component of weather-dependent decision-making, from everyday events with high socio-economic impact, such as aviation and open-air events, to extreme events related to disaster management. "]}
{"annotations": [{"id": 86962353, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86984385, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86984414, "completed_by": {"id": 126844}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_2", "final_parag": "The satellite-based part of the dataset contains one visible channel, namely C02 (central wavelength of 0.64 $\\mu m$ and a spatial resolution of 0.5 km at nadir), two infrared channels, namely C09 and C13 (central wavelengths of 6.9 $\\mu m$ and 10.3 $\\mu m$, respectively, and a spatial resolution of 2 km), and lightning counts (8 km when mapped into gridded pixels).", "candidate_comments": ["The dataset contains remote sensing data from two sensors aboard GOES-16: the Advanced Baseline Imager \\citep[ABI;][]{Schmit2018} and the Geostationary Lightning Mapper \\citep[GLM;][]{Goodman2013}. One visible and two infrared channels of the dataset are acquired by the ABI sensor, while the lightning counts are acquired by the GLM.  The visible channel (C02) has a central wavelength of 0.64 $\\mu m$ and a spatial resolution of 0.5 km at nadir. \nThe two infrared channels (C09 and C13) have central wavelengths of 6.9 $\\mu m$ and 10.3 $\\mu m$ and a spatial resolution of 2 km. The lightning counts channel has a spatial resolution of 8 km when mapped into gridded pixels."]}
{"annotations": [{"id": 86965241, "completed_by": {"id": 140156}, "result": ["Réécriture", "Réécriture"]}, {"id": 86983617, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent"]}, {"id": 86983854, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}], "agreement": 77.77777777777777, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_3", "final_parag": "To circumvent the problem of class imbalance during model training due to the prevalence of clear-sky conditions, \\citet{SEVIR} employed stratified sampling of random events, constituting approximately 80\\% of the dataset, based on VIL values within the domain.", "candidate_comments": ["Selecting patches from the space-time dimension is a critical step both to ensure representativeness of different conditions and to provide a meaningful handle for individual events during analysis.  To prevent clear-sky conditions from dominating the samples, \\citet{SEVIR} used stratified sampling of random events, which constitutes around 80 \\% of their dataset, based on the VIL values of the domain. ", "This approach mitigated the problem of class imbalance during model training by manipulating the weather conditions of the random events towards severe weather conditions. "]}
{"annotations": [{"id": 86721932, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86731505, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86729779, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_4", "final_parag": "Moreover, \\citet{Earthformer} introduced an efficient patching framework for three-dimensional data by computing attention on small cuboids to perform space-time prediction tasks from multiple domains, such as video prediction and nowcasting. The Earthformer architecture utilizes multiple self- and cross-attention blocks to attend the cuboids patched with different strategies, which are connected to each other with a ResNet-style architecture to utilize information available across a hierarchical encoder-decoder structure. The model learns statistical relationships within and between cuboids with self- and cross-attention, respectively.", "candidate_comments": ["While self-attention blocks learn statistical relationships within cuboids, cross-attention blocks  , where number of heads to attend can be configured  The model learns intra- and inter-cuboid relationships by attending to cuboids with self- and cross-attention blocks, respectively. \nThe model uses self- and cross-attention blocks to learn intra- and inter-cuboid relationships utilizes multiple self- and cross-attention blocks, where the model learns intra-cuboid relationships attends to cuboids "]}
{"annotations": [{"id": 87255510, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87258604, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87268288, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_5", "final_parag": "We refrained from using the visible channel to ensure comparable model performance for day and night conditions, despite its potential to improve model performance. \nWe normalized each channel by subtracting its mean value and dividing it by the standard deviation, both of which were computed over the training data. \nFinally, to ensure robust results during analysis, we considered samples from the second half of 2019 as the test data and used the remaining for training the model. Of the 11728 events remained after filtering them to have no missing data in the four channels used, 27 \\% of the events were used for testing while 70 and 3 \\% were used for training and validation, respectively.", "candidate_comments": ["\\subsubsection{Training details} The main objective of the model is to predict VIL image sequences up to 2 hours lead time at a spatial resolution of 1km using past image sequences of two infrared channels and lightning counts."]}
{"annotations": [{"id": 86961223, "completed_by": {"id": 62471}, "result": ["Réécriture", "Réécriture"]}, {"id": 86977293, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}, {"id": 86977686, "completed_by": {"id": 126844}, "result": ["Réécriture", "Différent"]}], "agreement": 77.77777777777777, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_6", "final_parag": "The objective of the model is to predict future time steps of VIL image sequences spanning over 120 minutes (24 frames) at a spatial resolution of 1km, given the past satellite image sequences of two infrared channels and lightning counts spanning over 125 minutes (25 frames).", "candidate_comments": ["\\subsubsection{Training details} The main objective of the model is to predict VIL image sequences up to 2 hours lead time at a spatial resolution of 1km using past image sequences of two infrared channels and lightning counts.", "Specifically, VIL image sequences composed of 24 frames in the future are predicted using satellite image sequences of 25 past data frames with 3 channels.  Specifically, VIL image sequences spanning to 120 minutes (24 frames) to the future are predicted using satellite image sequences with 3 channels from the past spanning to 125 minutes (25 frames).  To perform this task, several modifications were made to the data pipeline and to the Earthformer architecture. "]}
{"annotations": [{"id": 86864807, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86901590, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86902171, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_7", "final_parag": "We used the original loss function, mean squared error, proposed by \\citet{Earthformer}, but scaled the loss values across time dimension with an inverse linear function to give greater weight to earlier time steps of the model output.", "candidate_comments": ["Another important modification was to scale the loss values with respect to lead time. Specifically, we scaled the loss values of predictions using an inverse linear function to give greater weight to earlier time steps of the model output."]}
{"annotations": [{"id": 86899001, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86902197, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86909965, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_8", "final_parag": "Model performance for a single event is illustrated in Fig. \\ref{fig:modelOut}, where the input data are presented in the first three rows, the target in the fourth, and the predictions of our modified Earthformer model (EF Sat2Rad) and the 3D U-Net model as the baseline in the last two rows, respectively.", "candidate_comments": ["Model performance for a single event is depicted in Fig. \\ref{fig:modelOut}, where the input data are presented in the first three rows, the target in the fourth row, and the prediction in the last row. "]}
{"annotations": [{"id": 86983103, "completed_by": {"id": 140156}, "result": ["Réécriture", "Différent"]}, {"id": 86985629, "completed_by": {"id": 62306}, "result": ["Différent", "Réécriture"]}, {"id": 86985693, "completed_by": {"id": 70661}, "result": ["Différent", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_9", "final_parag": "Temperature variations in space are evident in both infrared channels of the input data, albeit sharper in C13, and it overlaps with lightning activity in the early stages of the convection.", "candidate_comments": ["Model performance for a single event is depicted in Fig. \\ref{fig:modelOut}, where the input data are presented in the first three rows, the target in the fourth row, and the prediction in the last row. ", "With respect to the input data, temperature change in space is visible in both of the infrared channels, albeit sharper in C13, and it overlaps with lightning activity in the early stages of the convection. "]}
{"annotations": [{"id": 87255564, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 87259850, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87268293, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_10", "final_parag": "Regarding the target VIL values, multiple storm cells with large VIL values have developed within the earlier lead times and later moved eastward.", "candidate_comments": ["As for the target VIL values to be predicted, the VIL field has already developed but keeps expanding with multiple spots having high intensity. "]}
{"annotations": [{"id": 86984835, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86986376, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86986205, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_11", "final_parag": "While both models show a tendency to smooth predicted fields with increasing lead time, performance of our model degrades slower than the baseline, indicating a better capture of temporal dependencies. Overall, given the problem's complexity, the high spatial resolution, and the extended lead time, our model successfully learned the statistical relationships between satellite retrievals and ground-based VIL mosaics. Moreover, our model effectively reproduced the fields' growth, decay, and temporal movement (see Appendix for additional examples).", "candidate_comments": ["However, model performance degrades with increasing lead time due to its tendency to smooth predicted fields.  While both models show a tendency to smooth predicted fields with increasing lead time, performance of our model degrades slower than the baseline, suggesting better capture of temporal dependencies."]}
{"annotations": [{"id": 87255646, "completed_by": {"id": 70661}, "result": ["Réécriture", "Réécriture"]}, {"id": 87259469, "completed_by": {"id": 126844}, "result": ["Réécriture", "Réécriture"]}, {"id": 87268302, "completed_by": {"id": 62306}, "result": ["Réécriture", "Réécriture"]}], "agreement": 100.0, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_12", "final_parag": "We quantified model performance over the entire test set using Fractional Skill Score \\citep[FSS;][]{Roberts2008} on full image patches with spatial scales of 5 and 11 km to consider neighboring pixels, with several thresholds ranging from 0.78 to 12.14 kg/m$^2$ of VIL, parallel to the ones used in \\citet{SEVIR}.", "candidate_comments": ["To quantify model performance over the entire test set, we used two evaluation metrics, namely the Fractional Skill Score \\citep[FSS;][]{Roberts2008} and the Structure-Amplitude-Location Score \\citep[SAL;][]{Wernli2008}. ", "FSS is a spatial verification technique that considers neighboring pixels when estimating the score for a given pixel, thus reduces double-penalty effect. In other words, FSS can answer a question like `How likely is the model to predict an event of a certain magnitude within a certain distance of an observed event?' and is well-suited to quantify spatially high-resolution forecasts \\citep{Brown2011}.  FSS is a spatial verification technique that reduces double-penalty effect by considers neighboring pixels when estimating the score for a given pixel, thus well-suited to quantify spatially high-resolution forecasts \\citep{Brown2011}. \nIn order to compute the FSS values, we selected several thresholds ranging from 0.78 to 12.14 kg/m$^2$ of VIL, parallel to the ones used in \\citet{SEVIR} that are based on important intensity levels \\citep{Robinson2001}, and used different spatial scales"]}
{"annotations": [{"id": 86970980, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86961786, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86965947, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_13", "final_parag": "To gain confidence in model predictions and diagnose importance of channels and temporal extent of the input data, we conducted two permutation tests by shuffling input of the test dataset across channel and time. We attributed the skill score, normalized reduction in FSS values computed with 5km spatial scale (summarized in Fig. \\ref{fig:modelPerf}), to importance of the permuted chunk.", "candidate_comments": ["In order to gain confidence in model predictions and diagnose important data sources, we performed a permutation test on the trained model. Specifically, we generated three additional test datasets by randomly shuffling one of the three channels of the input data, thus removing the relevant information available in that channel. Finally, we attributed the skill score, normalized reduction in FSS values (computed with 5 km spatial scale), to the feature importance of each channel. In other words, importance of a given channel $X$ is quantified with skill scores as $1 - FSS_{X}/FSS_{ref}$, where $FSS_{ref}$ is the original prediction score (summarized in Fig. \\ref{fig:modelPerf}a) and $FSS_{X}$ is the prediction score of the same input data with the $X$ channel permuted. "]}
{"annotations": [{"id": 86736285, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86953032, "completed_by": {"id": 140156}, "result": ["Réécriture"]}, {"id": 86958245, "completed_by": {"id": 62471}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_14", "final_parag": "This is unsurprising since C09 senses the mean temperature value of a given water column \\citep{Schmit2018} while the temperature profile within the column can vary significantly.", "candidate_comments": ["Since C09 is a channel that primarily targets the temperature of `midlevel water vapor', it senses the mean temperature value of a given water column \\citep{Schmit2018}. However, temperature profile within the column can vary significantly. Therefore, C09 may be less useful in capturing the temperature information of the vertical profile or other essential aspects of a given column, which explains the channel's low skill score for the model trained in this study. "]}
{"annotations": [{"id": 86904097, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86973284, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973608, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_15", "final_parag": "We gained insights from the model by performing permutation tests to infer importance of channel and length of the input data. We found that the C13 channel contains skillful information for all weather conditions, while the lightning count channel has the highest relative feature importance in severe weather conditions, particularly in shorter lead times. Furthermore, we found that the model's predictive skill is sensitive to input data length in nowcasting severe weather conditions. While permutation tests can introduce bias and should be interpreted cautiously, our findings corroborate the significance of both the C13 channel and lightning activity in severe weather nowcasting, thereby strengthening confidence in our analysis.", "candidate_comments": ["Nevertheless, our feature importance analysis confirms the importance of the C13 channel for satellite-based nowcasting in general, together with the significance of lightning activity and in nowcasting severe weather, which instills confidence in the analysis. \nNevertheless, our feature importance analysis corroborates with the literature on the importance of C13 channel for nowcasting, as well as lightning activity for nowcasting severe weather gives confidence to the analysis."]}
{"annotations": [{"id": 86903908, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86972993, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86973500, "completed_by": {"id": 70661}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_16", "final_parag": "Our Transformer-based model serves as a proof-of-concept, demonstrating the potential for satellite-based precipitation nowcasting over large domains without the necessity of explicit radar tower support, once it is trained.", "candidate_comments": ["Our Transformer-based model provides opportunities to support precipitation nowcasting over large domains without the need for explicit radar tower support, once trained.  Our Transformer-based model serves as a proof-of-concept to support precipitation nowcasting over large domains without the need for explicit radar tower support, once trained. "]}
{"annotations": [{"id": 87255678, "completed_by": {"id": 70661}, "result": ["Différent"]}, {"id": 87259304, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 87268045, "completed_by": {"id": 62306}, "result": ["Différent"]}], "agreement": 100.0, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_17", "final_parag": "Moreover, model output can be used as a radar proxy in regions with limited data availability.", "candidate_comments": ["Moreover, it can assist in the parameterization and data assimilation pipelines for NWP and hydrological models, and provide synthetic data for future studies that aim to enable synergistic usage of data from different sources. "]}
{"annotations": [{"id": 86862125, "completed_by": {"id": 70661}, "result": ["Différent", "Différent"]}, {"id": 86901721, "completed_by": {"id": 62306}, "result": ["Différent", "Différent"]}, {"id": 86901771, "completed_by": {"id": 62471}, "result": ["Différent", "Différent"]}], "agreement": 100.0, "source_file": "2310.19515-e10efd1456c47e9b.jsonl", "id_parag": "2310.19515-e10efd1456c47e9b_18", "final_parag": "Improving model performance with better downscaling of the input data and expanding the model with more input data remains to be addressed in future studies towards operational data-driven nowcasting.", "candidate_comments": ["Are any thoughts for evaluating the performance of the transformer with other models? Might need to elaborate more on this stratified random sampling in Discussions!", "performed well under different weather phenomena,  and offered valuable insights for radar nowcasting "]}
{"annotations": [{"id": 86960895, "completed_by": {"id": 62471}, "result": ["Réécriture"]}, {"id": 86973077, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974706, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 100.0, "source_file": "2310.19519-9e00701c815fb921.jsonl", "id_parag": "2310.19519-9e00701c815fb921_1", "final_parag": "Graphically, SCM $\\mathcal{M}$ is a Directed Acyclic Graph (DAG) of the tuple $<\\boldsymbol{U}, \\boldsymbol{V}, \\mathcal{F}, P(\\boldsymbol{U})>$. For $f_{V_i} \\in \\mathcal{F}$, a directed edge $(V_j \\rightarrow V_i)$ maps $V_j \\in \\boldsymbol{Pa}_{V_i}$ to $V_i$. The corresponding neural SCM is defined as $\\mathcal{M}(\\boldsymbol{\\theta}) \\triangleq <\\hat{\\boldsymbol{U}}, \\boldsymbol{V}, \\hat{\\mathcal{F}}, P(\\hat{\\boldsymbol{U}})>$, where structural functions $\\mathcal{F}$ are parameterized with $\\boldsymbol{\\theta} = \\{\\boldsymbol{\\theta}_{V_i}: V_i \\in \\boldsymbol{V}\\}$, each $\\hat{f}_{V_i}$ is a FeedForward Neural Network (FFN). $\\hat{\\boldsymbol{U}}$ over bi-directed edges in $\\mathcal{M}(\\theta)$ is equivalently transformed into uniform prior according to neural causal theories \\cite{xia2022neural}. With the language of $\\mathcal{M}(\\theta)$, counterfactual is formalized as a joint distribution over multiple interventions in $\\mathcal{M}(\\theta)$ \\cite{berrevoets2021learning}.", "candidate_comments": ["Graphically, SCM $\\mathcal{M}$ is a directed acyclic graph of the tuple $<\\boldsymbol{U}, \\boldsymbol{V}, \\mathcal{F}, P(\\boldsymbol{U})>$. For $f_{V_i} \\in \\mathcal{F}$, a directed edge $(V_j \\rightarrow V_i)$ maps $V_j \\in \\boldsymbol{Pa}_{V_i}$ to $V_i$. The corresponding neural SCM is defined as $\\mathcal{M}(\\boldsymbol{\\theta}) \\triangleq <\\hat{\\boldsymbol{U}}, \\boldsymbol{V}, \\hat{\\mathcal{F}}, P(\\hat{\\boldsymbol{U}})>$, where structural functions $\\mathcal{F}: \\boldsymbol{Pa}_{V_i} \\bigcup \\{\\hat{U}_i\\}\\rightarrow V_i$ are parameterized with $\\boldsymbol{\\theta} = \\{\\boldsymbol{\\theta}_{V_i}: V_i \\in \\boldsymbol{V}\\}$,  each $\\hat{f}_{V_i} \\in \\hat{\\mathcal{F}_{\\boldsymbol{\\theta}}}$ is a feedforward neural network. $\\hat{\\boldsymbol{U}}$ over bi-directed edges in $\\mathcal{M}$ used in neural causal models is equivalently transformed into samples from uniform distribution according to neural causal theories \\cite{xia2022neural}. With the language of neural causal model $\\mathcal{M}(\\theta)$, counterfactual is then formalized as a joint distribution of multiple interventions which may not simultaneously exist in $\\mathcal{M}(\\theta)$ \\cite{berrevoets2021learning}."]}
{"annotations": [{"id": 86861928, "completed_by": {"id": 70661}, "result": ["Réécriture"]}, {"id": 86961370, "completed_by": {"id": 126844}, "result": ["Réécriture"]}, {"id": 86959515, "completed_by": {"id": 62471}, "result": ["Différent"]}], "agreement": 66.66666666666666, "source_file": "2310.19519-9e00701c815fb921.jsonl", "id_parag": "2310.19519-9e00701c815fb921_2", "final_parag": "\\textit{Probability of Necessity}\\cite{pearl2000models} is formalized upon the counterfactual, which measures the necessity of current intervention to the observed results. In our task, the intervention is the recommendation policy and the result is the user feedback. \n[EQUATION]", "candidate_comments": ["With counterfactual defined, we can formalize a causal quantity which measure the necessity of current intervention to the observed results, \\ie \\textit{Probability of Necessity}\\cite{pearl2000models} ,"]}
{"annotations": [{"id": 86903999, "completed_by": {"id": 62471}, "result": ["Différent"]}, {"id": 86975345, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86975168, "completed_by": {"id": 62306}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19521-950d59ba3d79ddfa.jsonl", "id_parag": "2310.19521-950d59ba3d79ddfa_1", "final_parag": "\\begin{definition} A matrix ${\\bf A}\\in \\mathcal{Z}^{n\\times n}$ is called an M-matrix if ${\\bf A}$ is inverse-positive. That is ${\\bf A}^{-1}$ exists and each entry of ${\\bf A}^{-1}$ is nonnegative. \n\\end{definition}", "candidate_comments": ["\\begin{theorem} A matrix ${\\bf A}\\in \\mathcal{Z}^{n\\times n}$ is an M-matrix if and only if ${\\bf A}$ has all positive diagonal elements, and there exists a positive diagonal matrix ${\\bf D}$ such that ${\\bf A}{\\bf D}$ is strictly diagonally dominant. \n\\end{theorem} The following corollary is obtained when ${\\bf D}$ is the identity matrix:"]}
{"annotations": [{"id": 86971660, "completed_by": {"id": 126844}, "result": ["Différent"]}, {"id": 86973020, "completed_by": {"id": 62306}, "result": ["Réécriture"]}, {"id": 86974401, "completed_by": {"id": 140156}, "result": ["Réécriture"]}], "agreement": 66.66666666666666, "source_file": "2310.19521-950d59ba3d79ddfa.jsonl", "id_parag": "2310.19521-950d59ba3d79ddfa_2", "final_parag": "\\begin{theorem} A matrix ${\\bf A}\\in \\mathcal{Z}^{n\\times n}$ is an M-matrix if and only if ${\\bf A}$ is semi-positive. That is, there exists ${\\bf x}=(x_1,\\dots,x_n)^\\top$ with $x_i>0$ such that $({\\bf A}{\\bf x})_i > 0$ for all $1\\leq i\\leq n$. \n\\end{theorem}", "candidate_comments": ["\\begin{theorem} A matrix ${\\bf A}\\in \\mathcal{Z}^{n\\times n}$ is an M-matrix if and only if ${\\bf A}$ has all positive diagonal elements, and there exists a positive diagonal matrix ${\\bf D}$ such that ${\\bf A}{\\bf D}$ is strictly diagonally dominant. \n\\end{theorem} The following corollary is obtained when ${\\bf D}$ is the identity matrix:"]}