Upload 12 files

.gitignore

@@ -0,0 +1,134 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+pip-wheel-metadata/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+.python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+*.tsv
+*.jsonl
+*.zip
+output/

README.md

@@ -1,12 +1,2 @@
----
-title: Nlp4web
-emoji: 🦀
-colorFrom: yellow
-colorTo: yellow
-sdk: gradio
-sdk_version: 5.5.0
-app_file: app.py
-pinned: false
----
-
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# nlp4web
+Codebase of teaching materials for NLP4Web.

app.py

@@ -0,0 +1,733 @@
+# -*- coding: utf-8 -*-
+"""Kopie von HW1 (more instructed).ipynb
+
+Automatically generated by Colab.
+
+Original file is located at
+    https://colab.research.google.com/drive/1dGoZK5ZufqNgHm3hH8FEXe34rFqvwLOY
+"""
+from __future__ import annotations
+
+"""## Pre-requisite code
+
+The code within this section will be used in the tasks. Please do not change these code lines.
+
+### SciQ loading and counting
+"""
+
+from dataclasses import dataclass
+import pickle
+import os
+from typing import Iterable, Callable, List, Dict, Optional, Type, TypeVar
+from nlp4web_codebase.ir.data_loaders.dm import Document
+from collections import Counter
+import tqdm
+import re
+import nltk
+nltk.download("stopwords", quiet=True)
+from nltk.corpus import stopwords as nltk_stopwords
+
+LANGUAGE = "english"
+word_splitter = re.compile(r"(?u)\b\w\w+\b").findall
+stopwords = set(nltk_stopwords.words(LANGUAGE))
+
+
+def word_splitting(text: str) -> List[str]:
+    return word_splitter(text.lower())
+
+def lemmatization(words: List[str]) -> List[str]:
+    return words  # We ignore lemmatization here for simplicity
+
+def simple_tokenize(text: str) -> List[str]:
+    words = word_splitting(text)
+    tokenized = list(filter(lambda w: w not in stopwords, words))
+    tokenized = lemmatization(tokenized)
+    return tokenized
+
+T = TypeVar("T", bound="InvertedIndex")
+
+@dataclass
+class PostingList:
+    term: str  # The term
+    docid_postings: List[int]  # docid_postings[i] means the docid (int) of the i-th associated posting
+    tweight_postings: List[float]  # tweight_postings[i] means the term weight (float) of the i-th associated posting
+
+
+@dataclass
+class InvertedIndex:
+    posting_lists: List[PostingList]  # docid -> posting_list
+    vocab: Dict[str, int]
+    cid2docid: Dict[str, int]  # collection_id -> docid
+    collection_ids: List[str]  # docid -> collection_id
+    doc_texts: Optional[List[str]] = None  # docid -> document text
+
+    def save(self, output_dir: str) -> None:
+        os.makedirs(output_dir, exist_ok=True)
+        with open(os.path.join(output_dir, "index.pkl"), "wb") as f:
+            pickle.dump(self, f)
+
+    @classmethod
+    def from_saved(cls: Type[T], saved_dir: str) -> T:
+        index = cls(
+            posting_lists=[], vocab={}, cid2docid={}, collection_ids=[], doc_texts=None
+        )
+        with open(os.path.join(saved_dir, "index.pkl"), "rb") as f:
+            index = pickle.load(f)
+        return index
+
+
+# The output of the counting function:
+@dataclass
+class Counting:
+    posting_lists: List[PostingList]
+    vocab: Dict[str, int]
+    cid2docid: Dict[str, int]
+    collection_ids: List[str]
+    dfs: List[int]  # tid -> df
+    dls: List[int]  # docid -> doc length
+    avgdl: float
+    nterms: int
+    doc_texts: Optional[List[str]] = None
+
+def run_counting(
+    documents: Iterable[Document],
+    tokenize_fn: Callable[[str], List[str]] = simple_tokenize,
+    store_raw: bool = True,  # store the document text in doc_texts
+    ndocs: Optional[int] = None,
+    show_progress_bar: bool = True,
+) -> Counting:
+    """Counting TFs, DFs, doc_lengths, etc."""
+    posting_lists: List[PostingList] = []
+    vocab: Dict[str, int] = {}
+    cid2docid: Dict[str, int] = {}
+    collection_ids: List[str] = []
+    dfs: List[int] = []  # tid -> df
+    dls: List[int] = []  # docid -> doc length
+    nterms: int = 0
+    doc_texts: Optional[List[str]] = []
+    for doc in tqdm.tqdm(
+        documents,
+        desc="Counting",
+        total=ndocs,
+        disable=not show_progress_bar,
+    ):
+        if doc.collection_id in cid2docid:
+            continue
+        collection_ids.append(doc.collection_id)
+        docid = cid2docid.setdefault(doc.collection_id, len(cid2docid))
+        toks = tokenize_fn(doc.text)
+        tok2tf = Counter(toks)
+        dls.append(sum(tok2tf.values()))
+        for tok, tf in tok2tf.items():
+            nterms += tf
+            tid = vocab.get(tok, None)
+            if tid is None:
+                posting_lists.append(
+                    PostingList(term=tok, docid_postings=[], tweight_postings=[])
+                )
+                tid = vocab.setdefault(tok, len(vocab))
+            posting_lists[tid].docid_postings.append(docid)
+            posting_lists[tid].tweight_postings.append(tf)
+            if tid < len(dfs):
+                dfs[tid] += 1
+            else:
+                dfs.append(0)
+        if store_raw:
+            doc_texts.append(doc.text)
+        else:
+            doc_texts = None
+    return Counting(
+        posting_lists=posting_lists,
+        vocab=vocab,
+        cid2docid=cid2docid,
+        collection_ids=collection_ids,
+        dfs=dfs,
+        dls=dls,
+        avgdl=sum(dls) / len(dls),
+        nterms=nterms,
+        doc_texts=doc_texts,
+    )
+
+from nlp4web_codebase.ir.data_loaders.sciq import load_sciq
+sciq = load_sciq()
+counting = run_counting(documents=iter(sciq.corpus), ndocs=len(sciq.corpus))
+
+"""### BM25 Index"""
+
+
+from dataclasses import asdict, dataclass
+import math
+import os
+from typing import Iterable, List, Optional, Type
+import tqdm
+from nlp4web_codebase.ir.data_loaders.dm import Document
+
+
+@dataclass
+class BM25Index(InvertedIndex):
+
+    @staticmethod
+    def tokenize(text: str) -> List[str]:
+        return simple_tokenize(text)
+
+    @staticmethod
+    def cache_term_weights(
+        posting_lists: List[PostingList],
+        total_docs: int,
+        avgdl: float,
+        dfs: List[int],
+        dls: List[int],
+        k1: float,
+        b: float,
+    ) -> None:
+        """Compute term weights and caching"""
+
+        N = total_docs
+        for tid, posting_list in enumerate(
+            tqdm.tqdm(posting_lists, desc="Regularizing TFs")
+        ):
+            idf = BM25Index.calc_idf(df=dfs[tid], N=N)
+            for i in range(len(posting_list.docid_postings)):
+                docid = posting_list.docid_postings[i]
+                tf = posting_list.tweight_postings[i]
+                dl = dls[docid]
+                regularized_tf = BM25Index.calc_regularized_tf(
+                    tf=tf, dl=dl, avgdl=avgdl, k1=k1, b=b
+                )
+                posting_list.tweight_postings[i] = regularized_tf * idf
+
+    @staticmethod
+    def calc_regularized_tf(
+        tf: int, dl: float, avgdl: float, k1: float, b: float
+    ) -> float:
+        return tf / (tf + k1 * (1 - b + b * dl / avgdl))
+
+    @staticmethod
+    def calc_idf(df: int, N: int):
+        return math.log(1 + (N - df + 0.5) / (df + 0.5))
+
+    @classmethod
+    def build_from_documents(
+        cls: Type[BM25Index],
+        documents: Iterable[Document],
+        store_raw: bool = True,
+        output_dir: Optional[str] = None,
+        ndocs: Optional[int] = None,
+        show_progress_bar: bool = True,
+        k1: float = 0.9,
+        b: float = 0.4,
+    ) -> BM25Index:
+        # Counting TFs, DFs, doc_lengths, etc.:
+        counting = run_counting(
+            documents=documents,
+            tokenize_fn=BM25Index.tokenize,
+            store_raw=store_raw,
+            ndocs=ndocs,
+            show_progress_bar=show_progress_bar,
+        )
+
+        # Compute term weights and caching:
+        posting_lists = counting.posting_lists
+        total_docs = len(counting.cid2docid)
+        BM25Index.cache_term_weights(
+            posting_lists=posting_lists,
+            total_docs=total_docs,
+            avgdl=counting.avgdl,
+            dfs=counting.dfs,
+            dls=counting.dls,
+            k1=k1,
+            b=b,
+        )
+
+        # Assembly and save:
+        index = BM25Index(
+            posting_lists=posting_lists,
+            vocab=counting.vocab,
+            cid2docid=counting.cid2docid,
+            collection_ids=counting.collection_ids,
+            doc_texts=counting.doc_texts,
+        )
+        return index
+
+bm25_index = BM25Index.build_from_documents(
+    documents=iter(sciq.corpus),
+    ndocs=12160,
+    show_progress_bar=True,
+)
+bm25_index.save("output/bm25_index")
+
+"""### BM25 Retriever"""
+
+from nlp4web_codebase.ir.models import BaseRetriever
+from typing import Type
+from abc import abstractmethod
+
+
+class BaseInvertedIndexRetriever(BaseRetriever):
+
+    @property
+    @abstractmethod
+    def index_class(self) -> Type[InvertedIndex]:
+        pass
+
+    def __init__(self, index_dir: str) -> None:
+        self.index = self.index_class.from_saved(index_dir)
+
+    def get_term_weights(self, query: str, cid: str) -> Dict[str, float]:
+        toks = self.index.tokenize(query)
+        target_docid = self.index.cid2docid[cid]
+        term_weights = {}
+        for tok in toks:
+            if tok not in self.index.vocab:
+                continue
+            tid = self.index.vocab[tok]
+            posting_list = self.index.posting_lists[tid]
+            for docid, tweight in zip(
+                posting_list.docid_postings, posting_list.tweight_postings
+            ):
+                if docid == target_docid:
+                    term_weights[tok] = tweight
+                    break
+        return term_weights
+
+    def score(self, query: str, cid: str) -> float:
+        return sum(self.get_term_weights(query=query, cid=cid).values())
+
+    def retrieve(self, query: str, topk: int = 10) -> Dict[str, float]:
+        toks = self.index.tokenize(query)
+        docid2score: Dict[int, float] = {}
+        for tok in toks:
+            if tok not in self.index.vocab:
+                continue
+            tid = self.index.vocab[tok]
+            posting_list = self.index.posting_lists[tid]
+            for docid, tweight in zip(
+                posting_list.docid_postings, posting_list.tweight_postings
+            ):
+                docid2score.setdefault(docid, 0)
+                docid2score[docid] += tweight
+        docid2score = dict(
+            sorted(docid2score.items(), key=lambda pair: pair[1], reverse=True)[:topk]
+        )
+        return {
+            self.index.collection_ids[docid]: score
+            for docid, score in docid2score.items()
+        }
+
+
+class BM25Retriever(BaseInvertedIndexRetriever):
+
+    @property
+    def index_class(self) -> Type[BM25Index]:
+        return BM25Index
+
+bm25_retriever = BM25Retriever(index_dir="output/bm25_index")
+bm25_retriever.retrieve("What type of diseases occur when the immune system attacks normal body cells?")
+
+"""# TASK1: tune b and k1 (4 points)
+
+Tune b and k1 on the **dev** split of SciQ using the metric MAP@10. The evaluation function (`evalaute_map`) is provided. Record the values in `plots_k1` and `plots_b`. Do it in a greedy manner: as the influence from b is larger, please first tune b (with k1 fixed to the default value 0.9) and use the best value of b to further tune k1.
+
+$${\displaystyle {\text{score}}(D,Q)=\sum _{i=1}^{n}{\text{IDF}}(q_{i})\cdot {\frac {f(q_{i},D)\cdot (k_{1}+1)}{f(q_{i},D)+k_{1}\cdot \left(1-b+b\cdot {\frac {|D|}{\text{avgdl}}}\right)}}}$$
+"""
+
+from nlp4web_codebase.ir.data_loaders import Split
+import pytrec_eval
+
+
+def evaluate_map(rankings: Dict[str, Dict[str, float]], split=Split.dev) -> float:
+  metric = "map_cut_10"
+  qrels = sciq.get_qrels_dict(split)
+  evaluator = pytrec_eval.RelevanceEvaluator(sciq.get_qrels_dict(split), (metric,))
+  qps = evaluator.evaluate(rankings)
+  return float(np.mean([qp[metric] for qp in qps.values()]))
+
+"""Example of using the pre-requisite code:"""
+
+# Loading dataset:
+from nlp4web_codebase.ir.data_loaders.sciq import load_sciq
+sciq = load_sciq()
+counting = run_counting(documents=iter(sciq.corpus), ndocs=len(sciq.corpus))
+
+# Building BM25 index and save:
+bm25_index = BM25Index.build_from_documents(
+    documents=iter(sciq.corpus),
+    ndocs=12160,
+    show_progress_bar=True
+)
+bm25_index.save("output/bm25_index")
+
+# Loading index and use BM25 retriever to retrieve:
+bm25_retriever = BM25Retriever(index_dir="output/bm25_index")
+print(bm25_retriever.retrieve("What type of diseases occur when the immune system attacks normal body cells?"))  # the ranking
+
+plots_b: Dict[str, List[float]] = {
+    "X": [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0],
+    "Y": []
+}
+plots_k1: Dict[str, List[float]] = {
+    "X": [0, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0],
+    "Y": []
+}
+
+## YOUR_CODE_STARTS_HERE
+# Two steps should be involved:
+# Step 1. Fix k1 value to the default one 0.9,
+# go through all the candidate b values (0, 0.1, ..., 1.0),
+# and record in plots_b["Y"] the corresponding performances obtained via evaluate_map;
+# Step 2. Fix b to the best one in step 1. and do the same for k1.
+
+# Hint (on using the pre-requisite code):
+# - One can use the loaded sciq dataset directly (loaded in the pre-requisite code);
+# - One can build bm25_index with `BM25Index.build_from_documents`;
+# - One can use BM25Retriever to load the index and perform retrieval on the dev queries
+# (dev queries can be obtained via sciq.get_split_queries(Split.dev))
+
+import numpy as np
+
+for x in plots_b["X"]:
+  bm25_index = BM25Index.build_from_documents(
+      documents=iter(sciq.corpus),
+      ndocs=12160,
+      show_progress_bar=True,
+      k1=0.9,
+      b=x
+  )
+  bm25_index.save("output/bm25_index")
+  bm25_retriever = BM25Retriever(index_dir="output/bm25_index")
+  rankings = {}
+  for query in sciq.get_split_queries(Split.dev):
+    ranking = bm25_retriever.retrieve(query=query.text)
+    rankings[query.query_id] = ranking
+  result = evaluate_map(rankings, split=Split.dev)
+  plots_b["Y"].append(result)
+
+best_b = plots_b["X"][np.argmax(plots_b["Y"])]
+
+for x in plots_k1["X"]:
+  bm25_index = BM25Index.build_from_documents(
+      documents=iter(sciq.corpus),
+      ndocs=12160,
+      show_progress_bar=True,
+      k1=x,
+      b=best_b
+  )
+  bm25_index.save("output/bm25_index")
+  bm25_retriever = BM25Retriever(index_dir="output/bm25_index")
+  rankings = {}
+  for query in sciq.get_split_queries(Split.dev):
+    ranking = bm25_retriever.retrieve(query=query.text)
+    rankings[query.query_id] = ranking
+  result = evaluate_map(rankings, split=Split.dev)
+  plots_k1["Y"].append(result)
+
+"""Let's check the effectiveness gain on test after this tuning on dev"""
+
+default_map = 0.7849
+best_b = plots_b["X"][np.argmax(plots_b["Y"])]
+best_k1 = plots_k1["X"][np.argmax(plots_k1["Y"])]
+bm25_index = BM25Index.build_from_documents(
+    documents=iter(sciq.corpus),
+    ndocs=12160,
+    show_progress_bar=True,
+    k1=best_k1,
+    b=best_b
+)
+bm25_index.save("output/bm25_index")
+bm25_retriever = BM25Retriever(index_dir="output/bm25_index")
+rankings = {}
+for query in sciq.get_split_queries(Split.test):  # note this is now on test
+  ranking = bm25_retriever.retrieve(query=query.text)
+  rankings[query.query_id] = ranking
+optimized_map = evaluate_map(rankings, split=Split.test)  # note this is now on test
+
+"""# TASK2: CSC matrix and `CSCBM25Index` (12 points)
+
+Recall that we use Python lists to implement posting lists, mapping term IDs to the documents in which they appear. This is inefficient due to its naive design. Actually [Compressed Sparse Column matrix](https://docs.scipy.org/doc/scipy/reference/generated/scipy.sparse.csc_matrix.html) is very suitable for storing the posting lists and can boost the efficiency.
+
+## TASK2.1: learn about `scipy.sparse.csc_matrix` (2 point)
+
+Convert the matrix \begin{bmatrix}
+0 & 1 & 0 & 3 \\
+10 & 2 & 1 & 0 \\
+0 & 0 & 0 & 9
+\end{bmatrix} to a `csc_matrix` by specifying `data`, `indices`, `indptr` and `shape`.
+"""
+
+from scipy.sparse._csc import csc_matrix
+
+
+"""## TASK2.2: implement `CSCBM25Index` (4 points)
+
+Implement `CSCBM25Index` by completing the missing code. Note that `CSCInvertedIndex` is similar to `InvertedIndex` which we talked about during the class. The main difference is posting lists are represented by a CSC sparse matrix.
+"""
+
+@dataclass
+class CSCInvertedIndex:
+    posting_lists_matrix: csc_matrix  # docid -> posting_list
+    vocab: Dict[str, int]
+    cid2docid: Dict[str, int]  # collection_id -> docid
+    collection_ids: List[str]  # docid -> collection_id
+    doc_texts: Optional[List[str]] = None  # docid -> document text
+
+    def save(self, output_dir: str) -> None:
+        os.makedirs(output_dir, exist_ok=True)
+        with open(os.path.join(output_dir, "index.pkl"), "wb") as f:
+            pickle.dump(self, f)
+
+    @classmethod
+    def from_saved(cls: Type[T], saved_dir: str) -> T:
+        index = cls(
+            posting_lists_matrix=None, vocab={}, cid2docid={}, collection_ids=[], doc_texts=None
+        )
+        with open(os.path.join(saved_dir, "index.pkl"), "rb") as f:
+            index = pickle.load(f)
+        return index
+
+@dataclass
+class CSCBM25Index(CSCInvertedIndex):
+
+    @staticmethod
+    def tokenize(text: str) -> List[str]:
+        return simple_tokenize(text)
+
+    @staticmethod
+    def cache_term_weights(
+        posting_lists: List[PostingList],
+        total_docs: int,
+        avgdl: float,
+        dfs: List[int],
+        dls: List[int],
+        k1: float,
+        b: float,
+    ) -> csc_matrix:
+        """Compute term weights and caching"""
+
+        ## YOUR_CODE_STARTS_HERE
+        data = []
+        indices = []
+        indptr = [0]
+        count = 0
+        N = total_docs
+        print(N)
+        print(len(posting_lists))
+        for tid, posting_list in enumerate(
+            tqdm.tqdm(posting_lists, desc="Regularizing TFs")
+        ):
+            idf = CSCBM25Index.calc_idf(df=dfs[tid], N=N)
+            #print(len(posting_list.docid_postings))
+            for i in range(len(posting_list.docid_postings)):
+                docid = posting_list.docid_postings[i]
+                tf = posting_list.tweight_postings[i]
+                dl = dls[docid]
+                regularized_tf = CSCBM25Index.calc_regularized_tf(
+                    tf=tf, dl=dl, avgdl=avgdl, k1=k1, b=b
+                )
+                # Update the term weight with modified TF * modified IDF:
+                data.append(regularized_tf * idf)
+                #indices.append(docid)
+                indices.append(docid)
+                count = count + 1
+
+            indptr.append(count)
+        #shape = (len(posting_lists),len(posting_lists[0].docid_postings))
+        output_matrix = csc_matrix((data, indices, indptr),dtype=np.float32) #shape=(N, len(posting_lists)))
+        #csc_transpose = output_matrix.transpose()
+        #print(len(posting_lists))
+        print(output_matrix.shape)
+        print(count)
+        print(output_matrix.size)
+        return output_matrix
+        ## YOUR_CODE_ENDS_HERE
+
+    @staticmethod
+    def calc_regularized_tf(
+        tf: int, dl: float, avgdl: float, k1: float, b: float
+    ) -> float:
+        return tf / (tf + k1 * (1 - b + b * dl / avgdl))
+
+    @staticmethod
+    def calc_idf(df: int, N: int):
+        return math.log(1 + (N - df + 0.5) / (df + 0.5))
+
+    @classmethod
+    def build_from_documents(
+        cls: Type[CSCBM25Index],
+        documents: Iterable[Document],
+        store_raw: bool = True,
+        output_dir: Optional[str] = None,
+        ndocs: Optional[int] = None,
+        show_progress_bar: bool = True,
+        k1: float = 0.9,
+        b: float = 0.4,
+    ) -> CSCBM25Index:
+        # Counting TFs, DFs, doc_lengths, etc.:
+        counting = run_counting(
+            documents=documents,
+            tokenize_fn=CSCBM25Index.tokenize,
+            store_raw=store_raw,
+            ndocs=ndocs,
+            show_progress_bar=show_progress_bar,
+        )
+
+        # Compute term weights and caching:
+        posting_lists = counting.posting_lists
+        total_docs = len(counting.cid2docid)
+        posting_lists_matrix = CSCBM25Index.cache_term_weights(
+            posting_lists=posting_lists,
+            total_docs=total_docs,
+            avgdl=counting.avgdl,
+            dfs=counting.dfs,
+            dls=counting.dls,
+            k1=k1,
+            b=b,
+        )
+
+        # Assembly and save:
+        index = CSCBM25Index(
+            posting_lists_matrix=posting_lists_matrix,
+            vocab=counting.vocab,
+            cid2docid=counting.cid2docid,
+            collection_ids=counting.collection_ids,
+            doc_texts=counting.doc_texts,
+        )
+        return index
+
+csc_bm25_index = CSCBM25Index.build_from_documents(
+    documents=iter(sciq.corpus),
+    ndocs=12160,
+    show_progress_bar=True,
+    k1=best_k1,
+    b=best_b
+)
+csc_bm25_index.save("output/csc_bm25_index")
+
+
+class BaseCSCInvertedIndexRetriever(BaseRetriever):
+
+    @property
+    @abstractmethod
+    def index_class(self) -> Type[CSCInvertedIndex]:
+        pass
+
+    def __init__(self, index_dir: str) -> None:
+        self.index = self.index_class.from_saved(index_dir)
+
+    def get_term_weights(self, query: str, cid: str) -> Dict[str, float]:
+        ## YOUR_CODE_STARTS_HERE
+        toks = self.index.tokenize(query)
+        target_docid = self.index.cid2docid[cid]
+        term_weights = {}
+        matrix = self.index.posting_lists_matrix.astype(np.float64)
+        for tok in toks:
+            if tok not in self.index.vocab:
+                continue
+            tid = self.index.vocab[tok]
+            if matrix[target_docid, tid]!= 0:
+              term_weights[tok] = matrix[target_docid, tid]
+
+        return term_weights
+        ## YOUR_CODE_ENDS_HERE
+
+    def score(self, query: str, cid: str) -> float:
+        return sum(self.get_term_weights(query=query, cid=cid).values())
+
+    def retrieve(self, query: str, topk: int = 10) -> Dict[str, float]:
+        ## YOUR_CODE_STARTS_HERE
+        toks = self.index.tokenize(query)
+        docid2score: Dict[int, float] = {}
+        matrix = self.index.posting_lists_matrix.astype(np.float64)
+        for tok in toks:
+            if tok not in self.index.vocab:
+                continue
+            tid = self.index.vocab[tok]
+
+            #posting_list = self.index.posting_lists[tid]
+            #for i, docid in enumerate(posting_list.docid_postings):
+                #tweight = matrix[docid, i]
+                #docid2score.setdefault(docid, 0)
+                #docid2score[docid] += tweight
+
+            for docid in range(matrix.shape[0]):
+                tweight = matrix[docid, tid]
+                docid2score.setdefault(docid, 0)
+                docid2score[docid] += tweight
+
+        docid2score = dict(
+            sorted(docid2score.items(), key=lambda pair: pair[1], reverse=True)[:topk]
+        )
+        return {
+            self.index.collection_ids[docid]: score
+            for docid, score in docid2score.items()
+        }
+
+
+
+        ## YOUR_CODE_ENDS_HERE
+
+
+class CSCBM25Retriever(BaseCSCInvertedIndexRetriever):
+
+    @property
+    def index_class(self) -> Type[CSCBM25Index]:
+        return CSCBM25Index
+
+
+
+"""# TASK3: a search-engine demo based on Huggingface space (4 points)
+
+## TASK3.1: create the gradio app (2 point)
+
+Create a gradio app to demo the BM25 search engine index on SciQ. The app should have a single input variable for the query (of type `str`) and a single output variable for the returned ranking (of type `List[Hit]` in the code below). Please use the BM25 system with default k1 and b values.
+
+Hint: it should use a "search" function of signature:
+
+```python
+def search(query: str) -> List[Hit]:
+  ...
+```
+"""
+
+
+import gradio as gr
+from typing import TypedDict
+
+class Hit(TypedDict):
+  cid: str
+  score: float
+  text: str
+
+demo: Optional[gr.Interface] = None  # Assign your gradio demo to this variable
+return_type = List[Hit]
+
+## YOUR_CODE_STARTS_HERE
+bm25_index = BM25Index.build_from_documents(
+      documents=iter(sciq.corpus),
+      ndocs=12160,
+      show_progress_bar=True,
+  )
+bm25_index.save("output/bm25_index")
+bm25_retriever = BM25Retriever(index_dir="output/bm25_index")
+
+def search(query: str) -> List[Hit]:
+    l = []
+    for x,y in bm25_retriever.retrieve(query).items():
+        hit_object: Hit = {
+    "cid": x,
+    "score": y,
+    "text": sciq.corpus[bm25_retriever.index.cid2docid[x]]
+}
+        l.append(hit_object)
+    return l
+#print(search("What type of organism is commonly used in preparation of foods such as cheese and yogurt?"))
+demo = gr.Interface(
+    fn=search,
+    inputs="text",
+    outputs= "text",
+)
+## YOUR_CODE_ENDS_HERE
+demo.launch()
+
+
+
+

nlp4web_codebase/ir/analysis.py

@@ -0,0 +1,160 @@
+import os
+from typing import Dict, List, Optional, Protocol
+import pandas as pd
+import tqdm
+import ujson
+from nlp4web_codebase.ir.data_loaders import IRDataset
+
+
+def round_dict(obj: Dict[str, float], ndigits: int = 4) -> Dict[str, float]:
+    return {k: round(v, ndigits=ndigits) for k, v in obj.items()}
+
+
+def sort_dict(obj: Dict[str, float], reverse: bool = True) -> Dict[str, float]:
+    return dict(sorted(obj.items(), key=lambda pair: pair[1], reverse=reverse))
+
+
+def save_ranking_results(
+    output_dir: str,
+    query_ids: List[str],
+    rankings: List[Dict[str, float]],
+    query_performances_lists: List[Dict[str, float]],
+    cid2tweights_lists: Optional[List[Dict[str, Dict[str, float]]]] = None,
+):
+    os.makedirs(output_dir, exist_ok=True)
+    output_path = os.path.join(output_dir, "ranking_results.jsonl")
+    rows = []
+    for i, (query_id, ranking, query_performances) in enumerate(
+        zip(query_ids, rankings, query_performances_lists)
+    ):
+        row = {
+            "query_id": query_id,
+            "ranking": round_dict(ranking),
+            "query_performances": round_dict(query_performances),
+            "cid2tweights": {},
+        }
+        if cid2tweights_lists is not None:
+            row["cid2tweights"] = {
+                cid: round_dict(tws) for cid, tws in cid2tweights_lists[i].items()
+            }
+        rows.append(row)
+    pd.DataFrame(rows).to_json(
+        output_path,
+        orient="records",
+        lines=True,
+    )
+
+
+class TermWeightingFunction(Protocol):
+    def __call__(self, query: str, cid: str) -> Dict[str, float]: ...
+
+
+def compare(
+    dataset: IRDataset,
+    results_path1: str,
+    results_path2: str,
+    output_dir: str,
+    main_metric: str = "recip_rank",
+    system1: Optional[str] = None,
+    system2: Optional[str] = None,
+    term_weighting_fn1: Optional[TermWeightingFunction] = None,
+    term_weighting_fn2: Optional[TermWeightingFunction] = None,
+) -> None:
+    os.makedirs(output_dir, exist_ok=True)
+    df1 = pd.read_json(results_path1, orient="records", lines=True)
+    df2 = pd.read_json(results_path2, orient="records", lines=True)
+    assert len(df1) == len(df2)
+    all_qrels = {}
+    for split in dataset.split2qrels:
+        all_qrels.update(dataset.get_qrels_dict(split))
+    qid2query = {query.query_id: query for query in dataset.queries}
+    cid2doc = {doc.collection_id: doc for doc in dataset.corpus}
+    diff_col = f"{main_metric}:qp1-qp2"
+    merged = pd.merge(df1, df2, on="query_id", how="outer")
+    rows = []
+    for _, example in tqdm.tqdm(merged.iterrows(), desc="Comparing", total=len(merged)):
+        docs = {cid: cid2doc[cid].text for cid in dict(example["ranking_x"])}
+        docs.update({cid: cid2doc[cid].text for cid in dict(example["ranking_y"])})
+        query_id = example["query_id"]
+        row = {
+            "query_id": query_id,
+            "query": qid2query[query_id].text,
+            diff_col: example["query_performances_x"][main_metric]
+            - example["query_performances_y"][main_metric],
+            "ranking1": ujson.dumps(example["ranking_x"], indent=4),
+            "ranking2": ujson.dumps(example["ranking_y"], indent=4),
+            "docs": ujson.dumps(docs, indent=4),
+            "query_performances1": ujson.dumps(
+                example["query_performances_x"], indent=4
+            ),
+            "query_performances2": ujson.dumps(
+                example["query_performances_y"], indent=4
+            ),
+            "qrels": ujson.dumps(all_qrels[query_id], indent=4),
+        }
+        if term_weighting_fn1 is not None and term_weighting_fn2 is not None:
+            all_cids = set(example["ranking_x"]) | set(example["ranking_y"])
+            cid2tweights1 = {}
+            cid2tweights2 = {}
+            ranking1 = {}
+            ranking2 = {}
+            for cid in all_cids:
+                tweights1 = term_weighting_fn1(query=qid2query[query_id].text, cid=cid)
+                tweights2 = term_weighting_fn2(query=qid2query[query_id].text, cid=cid)
+                ranking1[cid] = sum(tweights1.values())
+                ranking2[cid] = sum(tweights2.values())
+                cid2tweights1[cid] = tweights1
+                cid2tweights2[cid] = tweights2
+            ranking1 = sort_dict(ranking1)
+            ranking2 = sort_dict(ranking2)
+            row["ranking1"] = ujson.dumps(ranking1, indent=4)
+            row["ranking2"] = ujson.dumps(ranking2, indent=4)
+            cid2tweights1 = {cid: cid2tweights1[cid] for cid in ranking1}
+            cid2tweights2 = {cid: cid2tweights2[cid] for cid in ranking2}
+            row["cid2tweights1"] = ujson.dumps(cid2tweights1, indent=4)
+            row["cid2tweights2"] = ujson.dumps(cid2tweights2, indent=4)
+        rows.append(row)
+    table = pd.DataFrame(rows).sort_values(by=diff_col, ascending=False)
+    output_path = os.path.join(output_dir, f"compare-{system1}_vs_{system2}.tsv")
+    table.to_csv(output_path, sep="\t", index=False)
+
+
+# if __name__ == "__main__":
+#     # python -m lecture2.bm25.analysis
+#     from nlp4web_codebase.ir.data_loaders.sciq import load_sciq
+#     from lecture2.bm25.bm25_retriever import BM25Retriever
+#     from lecture2.bm25.tfidf_retriever import TFIDFRetriever
+#     import numpy as np
+
+#     sciq = load_sciq()
+#     system1 = "bm25"
+#     system2 = "tfidf"
+#     results_path1 = f"output/sciq-{system1}/results/ranking_results.jsonl"
+#     results_path2 = f"output/sciq-{system2}/results/ranking_results.jsonl"
+#     index_dir1 = f"output/sciq-{system1}"
+#     index_dir2 = f"output/sciq-{system2}"
+#     compare(
+#         dataset=sciq,
+#         results_path1=results_path1,
+#         results_path2=results_path2,
+#         output_dir=f"output/sciq-{system1}_vs_{system2}",
+#         system1=system1,
+#         system2=system2,
+#         term_weighting_fn1=BM25Retriever(index_dir1).get_term_weights,
+#         term_weighting_fn2=TFIDFRetriever(index_dir2).get_term_weights,
+#     )
+
+#     # bias on #shared_terms of TFIDF:
+#     df1 = pd.read_json(results_path1, orient="records", lines=True)
+#     df2 = pd.read_json(results_path2, orient="records", lines=True)
+#     merged = pd.merge(df1, df2, on="query_id", how="outer")
+#     nterms1 = []
+#     nterms2 = []
+#     for _, row in merged.iterrows():
+#         nterms1.append(len(list(dict(row["cid2tweights_x"]).values())[0]))
+#         nterms2.append(len(list(dict(row["cid2tweights_y"]).values())[0]))
+#     percentiles = (5, 25, 50, 75, 95)
+#     print(system1, np.percentile(nterms1, percentiles), np.mean(nterms1).round(2))
+#     print(system2, np.percentile(nterms2, percentiles), np.mean(nterms2).round(2))
+#     # bm25 [ 3.  4.  5.  7. 11.] 5.64
+#     # tfidf [1. 2. 3. 5. 9.] 3.58

nlp4web_codebase/ir/data_loaders/__init__.py

@@ -0,0 +1,35 @@
+from dataclasses import dataclass
+from enum import Enum
+from typing import Dict, List
+from nlp4web_codebase.ir.data_loaders.dm import Document, Query, QRel
+
+
+class Split(str, Enum):
+    train = "train"
+    dev = "dev"
+    test = "test"
+
+
+@dataclass
+class IRDataset:
+    corpus: List[Document]
+    queries: List[Query]
+    split2qrels: Dict[Split, List[QRel]]
+
+    def get_stats(self) -> Dict[str, int]:
+        stats = {"|corpus|": len(self.corpus), "|queries|": len(self.queries)}
+        for split, qrels in self.split2qrels.items():
+            stats[f"|qrels-{split}|"] = len(qrels)
+        return stats
+
+    def get_qrels_dict(self, split: Split) -> Dict[str, Dict[str, int]]:
+        qrels_dict = {}
+        for qrel in self.split2qrels[split]:
+            qrels_dict.setdefault(qrel.query_id, {})
+            qrels_dict[qrel.query_id][qrel.collection_id] = qrel.relevance
+        return qrels_dict
+
+    def get_split_queries(self, split: Split) -> List[Query]:
+        qrels = self.split2qrels[split]
+        qids = {qrel.query_id for qrel in qrels}
+        return list(filter(lambda query: query.query_id in qids, self.queries))

nlp4web_codebase/ir/data_loaders/dm.py

@@ -0,0 +1,22 @@
+from dataclasses import dataclass
+from typing import Optional
+
+
+@dataclass
+class Document:
+    collection_id: str
+    text: str
+
+
+@dataclass
+class Query:
+    query_id: str
+    text: str
+
+
+@dataclass
+class QRel:
+    query_id: str
+    collection_id: str
+    relevance: int
+    answer: Optional[str] = None

nlp4web_codebase/ir/data_loaders/sciq.py

@@ -0,0 +1,86 @@
+from typing import Dict, List
+from nlp4web_codebase.ir.data_loaders import IRDataset, Split
+from nlp4web_codebase.ir.data_loaders.dm import Document, Query, QRel
+from datasets import load_dataset
+import joblib
+
+
+@(joblib.Memory(".cache").cache)
+def load_sciq(verbose: bool = False) -> IRDataset:
+    train = load_dataset("allenai/sciq", split="train")
+    validation = load_dataset("allenai/sciq", split="validation")
+    test = load_dataset("allenai/sciq", split="test")
+    data = {Split.train: train, Split.dev: validation, Split.test: test}
+
+    # Each duplicated record is the same to each other:
+    df = train.to_pandas() + validation.to_pandas() + test.to_pandas()
+    for question, group in df.groupby("question"):
+        assert len(set(group["support"].tolist())) == len(group)
+        assert len(set(group["correct_answer"].tolist())) == len(group)
+
+    # Build:
+    corpus = []
+    queries = []
+    split2qrels: Dict[str, List[dict]] = {}
+    question2id = {}
+    support2id = {}
+    for split, rows in data.items():
+        if verbose:
+            print(f"|raw_{split}|", len(rows))
+        split2qrels[split] = []
+        for i, row in enumerate(rows):
+            example_id = f"{split}-{i}"
+            support: str = row["support"]
+            if len(support.strip()) == 0:
+                continue
+            question = row["question"]
+            if len(support.strip()) == 0:
+                continue
+            if support in support2id:
+                continue
+            else:
+                support2id[support] = example_id
+            if question in question2id:
+                continue
+            else:
+                question2id[question] = example_id
+            doc = {"collection_id": example_id, "text": support}
+            query = {"query_id": example_id, "text": row["question"]}
+            qrel = {
+                "query_id": example_id,
+                "collection_id": example_id,
+                "relevance": 1,
+                "answer": row["correct_answer"],
+            }
+            corpus.append(Document(**doc))
+            queries.append(Query(**query))
+            split2qrels[split].append(QRel(**qrel))
+
+    # Assembly and return:
+    return IRDataset(corpus=corpus, queries=queries, split2qrels=split2qrels)
+
+
+if __name__ == "__main__":
+    # python -m nlp4web_codebase.ir.data_loaders.sciq
+    import ujson
+    import time
+
+    start = time.time()
+    dataset = load_sciq(verbose=True)
+    print(f"Loading costs: {time.time() - start}s")
+    print(ujson.dumps(dataset.get_stats(), indent=4))
+    # ________________________________________________________________________________
+    # [Memory] Calling __main__--home-kwang-research-nlp4web-ir-exercise-nlp4web-nlp4web-ir-data_loaders-sciq.load_sciq...
+    # load_sciq(verbose=True)
+    # |raw_train| 11679
+    # |raw_dev| 1000
+    # |raw_test| 1000
+    # ________________________________________________________load_sciq - 7.3s, 0.1min
+    # Loading costs: 7.260092735290527s
+    # {
+    #     "|corpus|": 12160,
+    #     "|queries|": 12160,
+    #     "|qrels-train|": 10409,
+    #     "|qrels-dev|": 875,
+    #     "|qrels-test|": 876
+    # }

nlp4web_codebase/ir/models/__init__.py

@@ -0,0 +1,21 @@
+from abc import ABC, abstractmethod
+from typing import Any, Dict, Type
+
+
+class BaseRetriever(ABC):
+
+    @property
+    @abstractmethod
+    def index_class(self) -> Type[Any]:
+        pass
+
+    def get_term_weights(self, query: str, cid: str) -> Dict[str, float]:
+        raise NotImplementedError
+
+    @abstractmethod
+    def score(self, query: str, cid: str) -> float:
+        pass
+
+    @abstractmethod
+    def retrieve(self, query: str, topk: int = 10) -> Dict[str, float]:
+        pass

requirements.txt

@@ -0,0 +1,10 @@
+nltk==3.8.1
+numpy==1.26.4
+scipy==1.13.1
+pandas==2.2.2
+tqdm==4.66.5
+ujson==5.10.0
+joblib==1.4.2
+datasets==3.0.1
+pytrec_eval==0.5
+gradio

setup.py

@@ -0,0 +1,37 @@
+from setuptools import setup, find_packages
+
+
+with open("README.md", "r", encoding="utf-8") as fh:
+    readme = fh.read()
+
+setup(
+    name="nlp4web-codebase",
+    version="0.0.0",
+    author="Kexin Wang",
+    author_email="[email protected]",
+    description="Codebase of teaching materials for NLP4Web.",
+    long_description=readme,
+    long_description_content_type="text/markdown",
+    url="https://https://github.com/kwang2049/nlp4web-codebase",
+    project_urls={
+        "Bug Tracker": "https://github.com/kwang2049/nlp4web-codebase/issues",
+    },
+    packages=find_packages(),
+    classifiers=[
+        "Programming Language :: Python :: 3",
+        "License :: OSI Approved :: Apache Software License",
+        "Operating System :: OS Independent",
+    ],
+    python_requires=">=3.10",
+    install_requires=[
+        "nltk==3.8.1",
+        "numpy==1.26.4",
+        "scipy==1.13.1",
+        "pandas==2.2.2",
+        "tqdm==4.66.5",
+        "ujson==5.10.0",
+        "joblib==1.4.2",
+        "datasets==3.0.1",
+        "pytrec_eval==0.5",
+    ],
+)