bge-base-financial-matryoshka / README.md

Chuangmail

Add new SentenceTransformer model.

b4fd8c5 verified 9 months ago

preview code

raw

history blame contribute delete

29.9 kB

metadata

language:
  - en
license: apache-2.0
tags:
  - sentence-transformers
  - sentence-similarity
  - feature-extraction
  - generated_from_trainer
  - dataset_size:6300
  - loss:MatryoshkaLoss
  - loss:MultipleNegativesRankingLoss
base_model: BAAI/bge-base-en-v1.5
widget:
  - source_sentence: >-
      What was Iron Mountain's physical records retention rate approximately 15
      years after entry into their facilities?
    sentences:
      - >-
        Garmin Connect and Garmin Connect Mobile are web and mobile platforms
        where users can track and analyze their fitness, activities and
        workouts, and wellness data.
      - >-
        More than 50% of physical records that entered Iron Mountain's
        facilities approximately 15 years ago are still there today.
      - >-
        In the first quarter of 2023, the divestiture of the company’s Longwall
        business was finalized, resulting in an unfavorable impact to operating
        profit of $586 million, primarily a non-cash item driven by the release
        of accumulated foreign currency translation.
  - source_sentence: How much did the company's currently payable U.S. taxes amount to in 2023?
    sentences:
      - In 2023, the currently payable U.S. taxes amounted to $2,705 million.
      - >-
        The Company expects to realize at least $500 million of incremental
        run-rate cost savings in addition to integration synergies.
      - >-
        During fiscal year 2023, we returned $210 million through our quarterly
        cash dividend program which was initiated in November 2020.
  - source_sentence: >-
      What was the percentage decline in GMS for the year ended December 31,
      2023 compared to 2022?
    sentences:
      - >-
        The Gross Merchandise Sales (GMS) decreased by 1.2% in 2023 compared to
        2022.
      - >-
        If, in the future, foreign exchange or capital control restrictions were
        to be imposed and become applicable to us, such restrictions could
        potentially reduce the amounts that we would be able to receive from our
        Macao, Hong Kong and mainland China subsidiaries.
      - >-
        Net cash provided by operating activities decreased by $2.0 billion in
        fiscal 2022 compared to fiscal 2021.
  - source_sentence: What was the operating income for the year 2023?
    sentences:
      - >-
        Effective January 1, 2021, CSC changed the designation of its corporate
        headquarters from San Francisco, California to Westlake, Texas.
      - The operating income for the year 2023 was reported as -$74.3 million.
      - >-
        Table 12 shows that the total risk-weighted assets under Basel 3 for
        credit risk at Bank of America amounted to $1,580 billion as of December
        31, 2023.
  - source_sentence: >-
      What was the total amount of tax incurred, collected, and remitted by AT&T
      in 2023?
    sentences:
      - >-
        For example, in response to regulatory developments in Europe, we
        announced plans to change the legal basis for behavioral advertising on
        Facebook and Instagram in the EU, EEA, and Switzerland from "legitimate
        interests" to "consent," and in November 2023 we began offering users in
        the region a "subscription for no ads" alternative.
      - >-
        Professional services expenses decreased $8 million in 2023 from 2022
        primarily due to lower consulting expenses related to bringing certain
        mortgage technology-related costs in-house, partially offset by higher
        legal expenses primarily related to the Black Knight acquisition.
      - >-
        Total taxes incurred, collected and remitted by AT&T during 2023 were
        $16,877.
pipeline_tag: sentence-similarity
library_name: sentence-transformers
metrics:
  - cosine_accuracy@1
  - cosine_accuracy@3
  - cosine_accuracy@5
  - cosine_accuracy@10
  - cosine_precision@1
  - cosine_precision@3
  - cosine_precision@5
  - cosine_precision@10
  - cosine_recall@1
  - cosine_recall@3
  - cosine_recall@5
  - cosine_recall@10
  - cosine_ndcg@10
  - cosine_mrr@10
  - cosine_map@100
model-index:
  - name: BGE base Financial Matryoshka
    results:
      - task:
          type: information-retrieval
          name: Information Retrieval
        dataset:
          name: dim 768
          type: dim_768
        metrics:
          - type: cosine_accuracy@1
            value: 0.6771428571428572
            name: Cosine Accuracy@1
          - type: cosine_accuracy@3
            value: 0.8328571428571429
            name: Cosine Accuracy@3
          - type: cosine_accuracy@5
            value: 0.8614285714285714
            name: Cosine Accuracy@5
          - type: cosine_accuracy@10
            value: 0.9085714285714286
            name: Cosine Accuracy@10
          - type: cosine_precision@1
            value: 0.6771428571428572
            name: Cosine Precision@1
          - type: cosine_precision@3
            value: 0.2776190476190476
            name: Cosine Precision@3
          - type: cosine_precision@5
            value: 0.17228571428571426
            name: Cosine Precision@5
          - type: cosine_precision@10
            value: 0.09085714285714284
            name: Cosine Precision@10
          - type: cosine_recall@1
            value: 0.6771428571428572
            name: Cosine Recall@1
          - type: cosine_recall@3
            value: 0.8328571428571429
            name: Cosine Recall@3
          - type: cosine_recall@5
            value: 0.8614285714285714
            name: Cosine Recall@5
          - type: cosine_recall@10
            value: 0.9085714285714286
            name: Cosine Recall@10
          - type: cosine_ndcg@10
            value: 0.7950953946105658
            name: Cosine Ndcg@10
          - type: cosine_mrr@10
            value: 0.7584574829931973
            name: Cosine Mrr@10
          - type: cosine_map@100
            value: 0.7618150097795325
            name: Cosine Map@100
      - task:
          type: information-retrieval
          name: Information Retrieval
        dataset:
          name: dim 512
          type: dim_512
        metrics:
          - type: cosine_accuracy@1
            value: 0.6785714285714286
            name: Cosine Accuracy@1
          - type: cosine_accuracy@3
            value: 0.8257142857142857
            name: Cosine Accuracy@3
          - type: cosine_accuracy@5
            value: 0.8642857142857143
            name: Cosine Accuracy@5
          - type: cosine_accuracy@10
            value: 0.9014285714285715
            name: Cosine Accuracy@10
          - type: cosine_precision@1
            value: 0.6785714285714286
            name: Cosine Precision@1
          - type: cosine_precision@3
            value: 0.2752380952380952
            name: Cosine Precision@3
          - type: cosine_precision@5
            value: 0.17285714285714282
            name: Cosine Precision@5
          - type: cosine_precision@10
            value: 0.09014285714285714
            name: Cosine Precision@10
          - type: cosine_recall@1
            value: 0.6785714285714286
            name: Cosine Recall@1
          - type: cosine_recall@3
            value: 0.8257142857142857
            name: Cosine Recall@3
          - type: cosine_recall@5
            value: 0.8642857142857143
            name: Cosine Recall@5
          - type: cosine_recall@10
            value: 0.9014285714285715
            name: Cosine Recall@10
          - type: cosine_ndcg@10
            value: 0.7927053640201507
            name: Cosine Ndcg@10
          - type: cosine_mrr@10
            value: 0.7574620181405893
            name: Cosine Mrr@10
          - type: cosine_map@100
            value: 0.7614007843308703
            name: Cosine Map@100
      - task:
          type: information-retrieval
          name: Information Retrieval
        dataset:
          name: dim 256
          type: dim_256
        metrics:
          - type: cosine_accuracy@1
            value: 0.68
            name: Cosine Accuracy@1
          - type: cosine_accuracy@3
            value: 0.81
            name: Cosine Accuracy@3
          - type: cosine_accuracy@5
            value: 0.8528571428571429
            name: Cosine Accuracy@5
          - type: cosine_accuracy@10
            value: 0.8971428571428571
            name: Cosine Accuracy@10
          - type: cosine_precision@1
            value: 0.68
            name: Cosine Precision@1
          - type: cosine_precision@3
            value: 0.27
            name: Cosine Precision@3
          - type: cosine_precision@5
            value: 0.17057142857142854
            name: Cosine Precision@5
          - type: cosine_precision@10
            value: 0.0897142857142857
            name: Cosine Precision@10
          - type: cosine_recall@1
            value: 0.68
            name: Cosine Recall@1
          - type: cosine_recall@3
            value: 0.81
            name: Cosine Recall@3
          - type: cosine_recall@5
            value: 0.8528571428571429
            name: Cosine Recall@5
          - type: cosine_recall@10
            value: 0.8971428571428571
            name: Cosine Recall@10
          - type: cosine_ndcg@10
            value: 0.7889658321825918
            name: Cosine Ndcg@10
          - type: cosine_mrr@10
            value: 0.7541865079365075
            name: Cosine Mrr@10
          - type: cosine_map@100
            value: 0.7582635867273656
            name: Cosine Map@100
      - task:
          type: information-retrieval
          name: Information Retrieval
        dataset:
          name: dim 128
          type: dim_128
        metrics:
          - type: cosine_accuracy@1
            value: 0.6614285714285715
            name: Cosine Accuracy@1
          - type: cosine_accuracy@3
            value: 0.8
            name: Cosine Accuracy@3
          - type: cosine_accuracy@5
            value: 0.8385714285714285
            name: Cosine Accuracy@5
          - type: cosine_accuracy@10
            value: 0.8914285714285715
            name: Cosine Accuracy@10
          - type: cosine_precision@1
            value: 0.6614285714285715
            name: Cosine Precision@1
          - type: cosine_precision@3
            value: 0.26666666666666666
            name: Cosine Precision@3
          - type: cosine_precision@5
            value: 0.16771428571428568
            name: Cosine Precision@5
          - type: cosine_precision@10
            value: 0.08914285714285713
            name: Cosine Precision@10
          - type: cosine_recall@1
            value: 0.6614285714285715
            name: Cosine Recall@1
          - type: cosine_recall@3
            value: 0.8
            name: Cosine Recall@3
          - type: cosine_recall@5
            value: 0.8385714285714285
            name: Cosine Recall@5
          - type: cosine_recall@10
            value: 0.8914285714285715
            name: Cosine Recall@10
          - type: cosine_ndcg@10
            value: 0.7751876221972102
            name: Cosine Ndcg@10
          - type: cosine_mrr@10
            value: 0.7381241496598633
            name: Cosine Mrr@10
          - type: cosine_map@100
            value: 0.7423110490736153
            name: Cosine Map@100
      - task:
          type: information-retrieval
          name: Information Retrieval
        dataset:
          name: dim 64
          type: dim_64
        metrics:
          - type: cosine_accuracy@1
            value: 0.6257142857142857
            name: Cosine Accuracy@1
          - type: cosine_accuracy@3
            value: 0.78
            name: Cosine Accuracy@3
          - type: cosine_accuracy@5
            value: 0.8214285714285714
            name: Cosine Accuracy@5
          - type: cosine_accuracy@10
            value: 0.8728571428571429
            name: Cosine Accuracy@10
          - type: cosine_precision@1
            value: 0.6257142857142857
            name: Cosine Precision@1
          - type: cosine_precision@3
            value: 0.26
            name: Cosine Precision@3
          - type: cosine_precision@5
            value: 0.16428571428571426
            name: Cosine Precision@5
          - type: cosine_precision@10
            value: 0.08728571428571427
            name: Cosine Precision@10
          - type: cosine_recall@1
            value: 0.6257142857142857
            name: Cosine Recall@1
          - type: cosine_recall@3
            value: 0.78
            name: Cosine Recall@3
          - type: cosine_recall@5
            value: 0.8214285714285714
            name: Cosine Recall@5
          - type: cosine_recall@10
            value: 0.8728571428571429
            name: Cosine Recall@10
          - type: cosine_ndcg@10
            value: 0.750742644383485
            name: Cosine Ndcg@10
          - type: cosine_mrr@10
            value: 0.7114563492063489
            name: Cosine Mrr@10
          - type: cosine_map@100
            value: 0.7163043069454876
            name: Cosine Map@100

BGE base Financial Matryoshka

This is a sentence-transformers model finetuned from BAAI/bge-base-en-v1.5 on the json dataset. It maps sentences & paragraphs to a 768-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Model Description

Model Type: Sentence Transformer
Base model: BAAI/bge-base-en-v1.5
Maximum Sequence Length: 512 tokens
Output Dimensionality: 768 tokens
Similarity Function: Cosine Similarity
Training Dataset:
- json
Language: en
License: apache-2.0

Model Sources

Documentation: Sentence Transformers Documentation
Repository: Sentence Transformers on GitHub
Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': True}) with Transformer model: BertModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
  (2): Normalize()
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("Chuangmail/bge-base-financial-matryoshka")
# Run inference
sentences = [
    'What was the total amount of tax incurred, collected, and remitted by AT&T in 2023?',
    'Total taxes incurred, collected and remitted by AT&T during 2023 were $16,877.',
    'Professional services expenses decreased $8 million in 2023 from 2022 primarily due to lower consulting expenses related to bringing certain mortgage technology-related costs in-house, partially offset by higher legal expenses primarily related to the Black Knight acquisition.',
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

Evaluation

Metrics

Information Retrieval

Dataset: dim_768
Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6771
cosine_accuracy@3	0.8329
cosine_accuracy@5	0.8614
cosine_accuracy@10	0.9086
cosine_precision@1	0.6771
cosine_precision@3	0.2776
cosine_precision@5	0.1723
cosine_precision@10	0.0909
cosine_recall@1	0.6771
cosine_recall@3	0.8329
cosine_recall@5	0.8614
cosine_recall@10	0.9086
cosine_ndcg@10	0.7951
cosine_mrr@10	0.7585
cosine_map@100	0.7618

Information Retrieval

Dataset: dim_512
Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6786
cosine_accuracy@3	0.8257
cosine_accuracy@5	0.8643
cosine_accuracy@10	0.9014
cosine_precision@1	0.6786
cosine_precision@3	0.2752
cosine_precision@5	0.1729
cosine_precision@10	0.0901
cosine_recall@1	0.6786
cosine_recall@3	0.8257
cosine_recall@5	0.8643
cosine_recall@10	0.9014
cosine_ndcg@10	0.7927
cosine_mrr@10	0.7575
cosine_map@100	0.7614

Information Retrieval

Dataset: dim_256
Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.68
cosine_accuracy@3	0.81
cosine_accuracy@5	0.8529
cosine_accuracy@10	0.8971
cosine_precision@1	0.68
cosine_precision@3	0.27
cosine_precision@5	0.1706
cosine_precision@10	0.0897
cosine_recall@1	0.68
cosine_recall@3	0.81
cosine_recall@5	0.8529
cosine_recall@10	0.8971
cosine_ndcg@10	0.789
cosine_mrr@10	0.7542
cosine_map@100	0.7583

Information Retrieval

Dataset: dim_128
Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6614
cosine_accuracy@3	0.8
cosine_accuracy@5	0.8386
cosine_accuracy@10	0.8914
cosine_precision@1	0.6614
cosine_precision@3	0.2667
cosine_precision@5	0.1677
cosine_precision@10	0.0891
cosine_recall@1	0.6614
cosine_recall@3	0.8
cosine_recall@5	0.8386
cosine_recall@10	0.8914
cosine_ndcg@10	0.7752
cosine_mrr@10	0.7381
cosine_map@100	0.7423

Information Retrieval

Dataset: dim_64
Evaluated with InformationRetrievalEvaluator

Metric	Value
cosine_accuracy@1	0.6257
cosine_accuracy@3	0.78
cosine_accuracy@5	0.8214
cosine_accuracy@10	0.8729
cosine_precision@1	0.6257
cosine_precision@3	0.26
cosine_precision@5	0.1643
cosine_precision@10	0.0873
cosine_recall@1	0.6257
cosine_recall@3	0.78
cosine_recall@5	0.8214
cosine_recall@10	0.8729
cosine_ndcg@10	0.7507
cosine_mrr@10	0.7115
cosine_map@100	0.7163

Training Details

Training Dataset

json

Dataset: json
Size: 6,300 training samples
Columns: anchor and positive
Approximate statistics based on the first 1000 samples:
anchor positive
type string string
details
min: 2 tokens
mean: 20.39 tokens
max: 40 tokens

min: 2 tokens
mean: 46.37 tokens
max: 326 tokens

	anchor	positive
type	string	string
details	min: 2 tokens mean: 20.39 tokens max: 40 tokens	min: 2 tokens mean: 46.37 tokens max: 326 tokens

Samples:

anchor	positive
`What are the key factors HP considers when making adjustments to inventory valuation?`	`HP makes adjustments to inventory valuation based on considerations of changes in demand, technological changes, supply constraints, product life cycle, component cost trends, product pricing, and quality issues.`
`What types of products does AbbVie's portfolio include?`	`AbbVie is a global, diversified research-based biopharmaceutical company with a comprehensive product portfolio that has leadership positions across immunology, oncology, aesthetics, neuroscience, and eye care.`
`What does IBM’s 2023 Annual Report to Stockholders include?`	`IBM's 2023 Annual Report to Stockholders includes their financial statements and supplementary data, which span from pages 44 to 121 and are incorporated by reference in the Form 10-K. Additionally, the financial statement schedule can be found on page S-1 of the same Form 10-K.`

Loss: MatryoshkaLoss with these parameters:

{
    "loss": "MultipleNegativesRankingLoss",
    "matryoshka_dims": [
        768,
        512,
        256,
        128,
        64
    ],
    "matryoshka_weights": [
        1,
        1,
        1,
        1,
        1
    ],
    "n_dims_per_step": -1
}

Training Hyperparameters

Non-Default Hyperparameters

eval_strategy: epoch
per_device_train_batch_size: 32
per_device_eval_batch_size: 16
gradient_accumulation_steps: 16
learning_rate: 2e-05
num_train_epochs: 4
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: True
tf32: True
load_best_model_at_end: True
optim: adamw_torch_fused
batch_sampler: no_duplicates

All Hyperparameters

Click to expand

overwrite_output_dir: False
do_predict: False
eval_strategy: epoch
prediction_loss_only: True
per_device_train_batch_size: 32
per_device_eval_batch_size: 16
per_gpu_train_batch_size: None
per_gpu_eval_batch_size: None
gradient_accumulation_steps: 16
eval_accumulation_steps: None
learning_rate: 2e-05
weight_decay: 0.0
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 1e-08
max_grad_norm: 1.0
num_train_epochs: 4
max_steps: -1
lr_scheduler_type: cosine
lr_scheduler_kwargs: {}
warmup_ratio: 0.1
warmup_steps: 0
log_level: passive
log_level_replica: warning
log_on_each_node: True
logging_nan_inf_filter: True
save_safetensors: True
save_on_each_node: False
save_only_model: False
restore_callback_states_from_checkpoint: False
no_cuda: False
use_cpu: False
use_mps_device: False
seed: 42
data_seed: None
jit_mode_eval: False
use_ipex: False
bf16: True
fp16: False
fp16_opt_level: O1
half_precision_backend: auto
bf16_full_eval: False
fp16_full_eval: False
tf32: True
local_rank: 0
ddp_backend: None
tpu_num_cores: None
tpu_metrics_debug: False
debug: []
dataloader_drop_last: False
dataloader_num_workers: 0
dataloader_prefetch_factor: None
past_index: -1
disable_tqdm: False
remove_unused_columns: True
label_names: None
load_best_model_at_end: True
ignore_data_skip: False
fsdp: []
fsdp_min_num_params: 0
fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
fsdp_transformer_layer_cls_to_wrap: None
accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
deepspeed: None
label_smoothing_factor: 0.0
optim: adamw_torch_fused
optim_args: None
adafactor: False
group_by_length: False
length_column_name: length
ddp_find_unused_parameters: None
ddp_bucket_cap_mb: None
ddp_broadcast_buffers: False
dataloader_pin_memory: True
dataloader_persistent_workers: False
skip_memory_metrics: True
use_legacy_prediction_loop: False
push_to_hub: False
resume_from_checkpoint: None
hub_model_id: None
hub_strategy: every_save
hub_private_repo: False
hub_always_push: False
gradient_checkpointing: False
gradient_checkpointing_kwargs: None
include_inputs_for_metrics: False
eval_do_concat_batches: True
fp16_backend: auto
push_to_hub_model_id: None
push_to_hub_organization: None
mp_parameters:
auto_find_batch_size: False
full_determinism: False
torchdynamo: None
ray_scope: last
ddp_timeout: 1800
torch_compile: False
torch_compile_backend: None
torch_compile_mode: None
dispatch_batches: None
split_batches: None
include_tokens_per_second: False
include_num_input_tokens_seen: False
neftune_noise_alpha: None
optim_target_modules: None
batch_eval_metrics: False
batch_sampler: no_duplicates
multi_dataset_batch_sampler: proportional

Training Logs

Epoch	Step	Training Loss	dim_128_cosine_map@100	dim_256_cosine_map@100	dim_512_cosine_map@100	dim_64_cosine_map@100	dim_768_cosine_map@100
0.8122	10	1.6191	-	-	-	-	-
0.9746	12	-	0.7267	0.7355	0.7447	0.6939	0.7453
1.6244	20	0.6415	-	-	-	-	-
1.9492	24	-	0.7358	0.7509	0.7548	0.7075	0.7554
2.4365	30	0.4638	-	-	-	-	-
2.9239	36	-	0.7398	0.7573	0.7607	0.7124	0.7601
3.2487	40	0.4083	-	-	-	-	-
3.8985	48	-	0.7423	0.7583	0.7614	0.7163	0.7618

The bold row denotes the saved checkpoint.

Framework Versions

Python: 3.11.11
Sentence Transformers: 3.1.0
Transformers: 4.41.2
PyTorch: 2.1.2+cu121
Accelerate: 1.3.0
Datasets: 2.19.1
Tokenizers: 0.19.1

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

MatryoshkaLoss

@misc{kusupati2024matryoshka,
    title={Matryoshka Representation Learning},
    author={Aditya Kusupati and Gantavya Bhatt and Aniket Rege and Matthew Wallingford and Aditya Sinha and Vivek Ramanujan and William Howard-Snyder and Kaifeng Chen and Sham Kakade and Prateek Jain and Ali Farhadi},
    year={2024},
    eprint={2205.13147},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}

MultipleNegativesRankingLoss

@misc{henderson2017efficient,
    title={Efficient Natural Language Response Suggestion for Smart Reply},
    author={Matthew Henderson and Rami Al-Rfou and Brian Strope and Yun-hsuan Sung and Laszlo Lukacs and Ruiqi Guo and Sanjiv Kumar and Balint Miklos and Ray Kurzweil},
    year={2017},
    eprint={1705.00652},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}