AwesomeMRC

update

• [model] Our SOTA SQuAD2.0 models are available at CodaLab for reproduction.

• [ensemble] The code for model ensemble (grid search): run_ensemble_grid.py

Requirements

The codes are based on Transformers v2.3.0. The dependencies are the same. You can install the dependencies by pip install transformers==2.3.0

or directly download the requirements file: https://github.com/huggingface/transformers/blob/v2.3.0/requirements.txt and run pip install -r requirements.

Summary

Looking for a comprehensive and comparative review of MRC? check out our new survey paper: Machine Reading Comprehension: The Role of Contextualized Language Models and Beyond (preprint, 2020).

In this work, MRC model is regarded as a two-stage Encoder-Decoder architecture. Our empirical analysis is shared in this repo.

Encoder:

1. Language Units

   Subword-augmented Embedding for Cloze Reading Comprehension (COLING 2018)

   Effective Subword Segmentation for Text Comprehension (TASLP)

2. Linguistic Knowledge

   Semantics-aware BERT for language understanding (AAAI 2020)

   SG-Net: Syntax-Guided Machine Reading Comprehension (AAAI 2020)

   LIMIT-BERT: Linguistic Informed Multi-Task BERT (preprint)

3. Commonsense Injection

   Multi-choice Dialogue-Based Reading Comprehension with Knowledge and Key Turns (preprint)

4. Contextualized language models (CLMs) for MRC:

Decoder:

The implementation is based on Transformers v2.3.0.

As part of the techniques in our Retro-Reader paper:

Retrospective Reader for Machine Reading Comprehension (AAAI 2021)

Answer Verification

1) Multitask-style verification

We evaluate different loss functions

cross-entropy (run_squad_av.py)

binary cross-entropy (run_squad_av_bce.py)

mse regression (run_squad_avreg.py)

2) External verification

Train an external verifier (run_cls.py)

Matching Network

Cross Attention (run_squad_seq_trm.py)

Matching Attention (run_squad_seq_sc.py)

Related Work:

Modeling Multi-turn Conversation with Deep Utterance Aggregation (COLING 2018)

DCMN+: Dual Co-Matching Network for Multi-choice Reading Comprehension (AAAI 2020)

Answer Dependency

Model answer dependency (start + seq -> end) (run_squad_dep.py)

Example: Retrospective Reader

1. train a sketchy reader (sh_albert_cls.sh)

2. train an intensive reader (sh_albert_av.sh)

3. rear verification: merge the prediction for final answer (run_verifier.py)

SQuAD 2.0 Dev Results:	

  ```
  {
  "exact": 87.75372694348522, 
  "f1": 90.91630165754992, 
  "total": 11873, 
  "HasAns_exact": 83.1140350877193, 
  "HasAns_f1": 89.4482539777485, 
  "HasAns_total": 5928, 
  "NoAns_exact": 92.38015138772077, 
  "NoAns_f1": 92.38015138772077, 
  "NoAns_total": 5945
  }
  ```

Question Classification

One-shot Learning for Question-Answering in Gaokao History Challenge (COLING 2018)

Citation

@article{zhang2020survey,
  title={Machine Reading Comprehension: The Role of Contextualized Language Models and Beyond},
  author={Zhang, Zhuosheng and Zhao, Hai and Wang, Rui},
  journal={arXiv preprint arXiv:2005.06249},
  year={2020}
}
@inproceedings{zhang2021retrospective,
  title={Retrospective reader for machine reading comprehension},
  author={Zhang, Zhuosheng and Yang, Junjie and Zhao, Hai},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={35},
  number={16},
  pages={14506--14514},
  year={2021}
}

Related Records (best)

CMRC 2017: The best single model (2017).

SQuAD 2.0: The best among all submissions (both single and ensemble settings); The first to surpass human benchmark on both EM and F1 scores with a single model (2019).

SNLI: The best among all submissions (2019-2020).

RACE: The best among all submissions (2019).

GLUE: The 3rd best among all submissions (early 2019).

Contact

Feel free to email zhangzs [at] sjtu.edu.cn if you have any questions.