Codes for coreference-aware machine reading comprehension

Data and code for the paper "Tracing Origins: Coreference-aware Machine Reading Comprehension" at ACL2022.

Dataset

There are three folders for our three models mentioned in the paper: Coref_additive_spacy for Coref_additive_attention, Coref_dgl_spacy for GNN and Coref_multiplication_spacy for Coref_multiplication_attention, and each contains the train data set and the dev data set under the quoref folder.

each sample contains

• context: the paragraph text
• context_id: the unique identifier of the context
• qas: a group of questions
• question: question text
• id: the unique identifier of the question
• answers: a group of the answers to one question
• text: answer text
• answer_start: the start_position of one answer

Models

If you want to use our trained model, please download it from Google drive

Training

python run_quoref.py --train_file "quoref/train.json" --predict_file "quoref/dev.json" --model_type "roberta_multi" --model_name_or_path "roberta-large" --output_dir "out" --do_train --do_eval --eval_all_checkpoints --learning_rate 1e-5 --num_train_epochs 6 --overwrite_output_dir --per_gpu_train_batch_size 4 --save_steps 6000 --coref_weight 0.4

Kindly Hint

There is an open issue regarding the compatibility between NeuralCoref and spaCy 3.0. If you intend to use the latest spaCy models, please watch the issue.

Cite

If you extend or use this work, please cite the paper where it was introduced:

@article{Huang2021TracingOC,
  title={Tracing Origins: Coref-aware Machine Reading Comprehension},
  author={Baorong Huang and Zhuosheng Zhang and Hai Zhao},
  journal={ArXiv},
  year={2021},
  volume={abs/2110.07961}
}