RCM-Question-Answering

This is a re-implementation of RCM-BERT question answering model by the ACL 2020 paper "Recurrent Chunking Mechanisms for Long-Text Machine Reading Comprehension".

Required package:

(1) python3;

(2) torch;

(3) Transformers by Huggingface.

Data Preparation

1. CoQA

(1) Download data from CoQA website and save to DATA_DIR.

Enter the directory RCM-Question-Answering/src/.

(2) Preprocess CoQA data

Start a CoreNLP server

  mkdir lib
  wget -O lib/stanford-corenlp-3.9.1.jar https://search.maven.org/remotecontent?filepath=edu/stanford/nlp/stanford-corenlp/3.9.1/stanford-corenlp-3.9.1.jar
  java -mx4g -cp lib/stanford-corenlp-3.9.1.jar edu.stanford.nlp.pipeline.StanfordCoreNLPServer -port 9000 -timeout 15000

Run a script to preprocess data

python3 data_helper/preprocess_coqa.py --data_file DATA_DIR/coqa-train-v1.0.json --output_file DATA_DIR/coqa.train.json
python3 data_helper/preprocess_coqa.py --data_file DATA_DIR/coqa-dev-v1.0.json --output_file DATA_DIR/coqa.dev.json

2. QuAC

Download data from QuAC websiet.

3. TriviaQA

Download data from TriviaQA websiet and save in DATA_DIR/. Two folders qa/ and evidence/ can be found in DATA_DIR/. TriviaQA dataset contains data from two domains: web and wikipedia, and we use its wikipedia portion in our experiments.

Follow the instructions here, and refer to its script utils.convert_to_squad_format.py to adapt TriviaQA to SQuAD format. Create a subfolder under squad-qa/ under the data folder DATA_DIR.

Clone the Github repo triviaqa, and convert data to squad format in squad-qa/

python3 -m utils.convert_to_squad_format --triviaqa_file DATA_DIR/qa/wikipedia-train.json --squad_file DATA_DIR/squad-qa/wikipedia-train.json --wikipedia_dir DATA_DIR/evidence/wikipedia/ --web_dir DATA_DIR/evidence/web/ --tokenizer NLTK_TOKENIZER_PATH [~/nltk_data/tokenizers/punkt/english.pickle]

python3 utils.convert_to_squad_format --triviaqa_file DATA_DIR/qa/wikipedia-dev.json --squad_file DATA_DIR/squad-qa/wikipedia-dev.json --wikipedia_dir DATA_DIR/evidence/wikipedia/ --web_dir DATA_DIR/evidence/web/ --tokenizer NLTK_TOKENIZER_PATH

• NLTK_TOKENIZER_PATH: path to nltk English tokenizer (tokenizers/punkt/english.pickle)

Training

For the efficiency of model training, we try first-pretrain-then-train approach. The model is first pre-trained with fixed strides and no recurrence. Then the recurrent chunking mechanism is applied to further train the model to chunk documents with flexible strides and progapagte informaiton among segmenets with recurrence.

1. Conversational Question Answering (CoQA)

Pretrain CoQA model

python3 train/run_BERT_coqa.py
--bert_model bert-large-uncased
--output_dir OUTPUT_DIR/pretrained/
--train_file DATA_DIR/coqa.train.josn
--use_history
--n_history -1
--max_seq_length MAX_SEQ_LENGTH
--doc_stride DOC_STRIDE
--max_query_length MAX_QUERY_LENGTH
--do_train
--do_validate
--train_batch_size 12
--predict_batch_size 18
--learning_rate 3e-5
--num_train_epochs 2.5
--max_answer_length 30
--do_lower_case

• MAX_SEQ_LENGTH: the maximum length of input sequence (<= 512).

• DOC_STRIDE: the stride size when reading through a document (set as 64 in the experiments).

• MAX_QUERY_LENGTH: the maximum length of a query in an input sequence (set as 64 in the experiments).

Recurrent chunking mechamism (RCM) for CoQA.

python3 train/run_RCM_coqa.py 
--bert_model bert-large-uncased 
--output_dir OUTPUT_DIR/rl/
--train_file DATA_DIR/coqa.train.json
--use_history
--n_history -1
--max_seq_length MAX_SEQ_LENGTH
--max_query_length MAX_QUERY_LENGTH
--do_train
--do_validate
--do_lower_case
--pretrained_model_path OUTPUT_DIR/pretrained/best_RCM_model.bin
--recur_type RECUR_TYPE
--train_batch_size 8
--learning_rate 1e-5
--num_train_epochs 2
--max_read_times MAX_READ_TIMES
--max_answer_length 30

• OUTPUT_DIR: the directory where the model and predictions are saved.

• DATA_DIR: the directory where the CoQA data is saved.

• MAX_SEQ_LENGTH: input sequence length and is larger than 512.

• MAX_QUERY_LENGTH: the maximum length of a query in an input sequence (set as 64).

• RECUR_TYPE: can be "gated" or "lstm" corresponding to Gated and LSTM recurrence respectively.

• MAX_READ_TIMES: the maximum number of segments read from a document. It is set as 4, 3, 3, 2 when the MAX_SEQ_LENGTH is set as 192, 256, 384 and 512 respectively.

Prediction

python3 train/run_RCM_coqa.py 
--bert_model bert-large-uncased 
--output_dir OUTPUT_DIR/rl/
--predict_file DATA_DIR/coqa.dev.json
--use_history
--n_history -1
--max_seq_length MAX_SEQ_LENGTH
--max_query_length MAX_QUERY_LENGTH
--doc_stride DOC_STRIDE
--do_predict
--do_lower_case
--recur_type RECUR_TYPE
--predict_batch_size 12
--max_read_times MAX_READ_TIMES
--max_answer_length 30

• Predictions will be saved in OUTPUT_DIR/rl/predictions.json

Evaluation

Download official evaluation script and save it in the folder evaluation/

python -m evaluation.evaluate-v1.0 --data-file DATA_DIR/coqa-dev-v1.0.json --pred-file OUTPUT_DIR/rl/predictions.json

2. Question Answering in Context (QuAC)

Pretrain QuAC model

python3 train/run_BERT_quac.py
--bert_model bert-large-uncased
--output_dir OUTPUT_DIR/pretrained/
--train_file DATA_DIR/train_v0.2.json
--use_history
--n_history -1
--max_seq_length MAX_SEQ_LENGTH
--doc_stride DOC_STRIDE
--max_query_length MAX_QUERY_LENGTH
--do_train
--do_validate
--train_batch_size 12
--predict_batch_size 18
--learning_rate 3e-5
--num_train_epochs 2.5
--max_answer_length 40
--do_lower_case

• MAX_SEQ_LENGTH: the maximum length of input sequence (<= 512).

• DOC_STRIDE: the stride size when reading through a document (set as 64 in the experiments).

• MAX_QUERY_LENGTH: the maximum length of a query in an input sequence (set as 64 in the experiments).

Recurrent chunking mechamism (RCM) for QuAC

python3 train/run_RCM_quac.py 
--bert_model bert-large-uncased 
--output_dir OUTPUT_DIR/rl/
--train_file DATA_DIR/train_v0.2.json
--use_history
--n_history -1
--max_seq_length MAX_SEQ_LENGTH
--max_query_length 64
--do_train
--do_validate
--do_lower_case
--pretrained_model_path OUTPUT_DIR/pretrained/best_RCM_model.bin
--recur_type RECUR_TYPE
--train_batch_size 8
--learning_rate 1e-5
--num_train_epochs 2.0
--max_read_times MAX_READ_TIMES
--max_answer_length 40

• MAX_SEQ_LENGTH: input sequence length and is larger than 512.

• MAX_QUERY_LENGTH: the maximum length of a query in an input sequence (set as 64).

• RECUR_TYPE: can be "gated" or "lstm" corresponding to Gated and LSTM recurrence respectively.

• MAX_READ_TIMES: the maximum number of segments read from a document. It is set as 4, 3, 3, 2 when the MAX_SEQ_LENGTH is set as 192, 256, 384 and 512 respectively.

Prediction

python3 train/run_RCM_quac.py 
--bert_model bert-large-uncased 
--output_dir OUTPUT_DIR/rl/
--predict_file DATA_DIR/val_v0.2.json
--use_history
--n_history -1
--max_seq_length MAX_SEQ_LENGTH
--max_query_length MAX_QUERY_LENGTH
--doc_stride DOC_STRIDE
--do_predict
--do_lower_case
--recur_type RECUR_TYPE
--predict_batch_size 12
--max_read_times MAX_READ_TIMES
--max_answer_length 40

• Predictions will be saved in OUTPUT_DIR/rl/predictions.json

Evaluation

The official evaluation script can be downloaded here. Since the script not only evaluates answer span predictions but also yes-no and followup predictions, we modify the evaluation script by only keeping its evaluation of predicted answer spans. The modified scripts can be found in evaluation/quac_evaluation

python -m evaluation.quac_evaluation --val_file DATA_DIR/val_v0.2.json --model_output OUTPUT_DIR/rl/predictions.json

3. TriviaQA

Pretrain Trivia model

python3 train/run_BERT_trivia.py
--bert_model bert-large-uncased
--output_dir OUTPUT_DIR/pretrained/
--train_file DATA_DIR/squad-qa/wikipedia-train.josn
--max_seq_length MAX_SEQ_LENGTH
--doc_stride DOC_STRIDE
--max_query_length MAX_QUERY_LENGTH
--do_train
--do_validate
--train_batch_size 12
--predict_batch_size 18
--learning_rate 3e-5
--num_train_epochs 2.5
--max_answer_length 60
--do_lower_case

• MAX_SEQ_LENGTH: the maximum length of input sequence, set as 512 in the experiment.

• DOC_STRIDE: the stride size when reading through a document.

• MAX_QUERY_LENGTH: the maximum length of a query in an input sequence (set as 64 in the experiments).

Recurrent chunking mechamism (RCM) for Trivia

python3 train/run_RCM_trivia.py 
--bert_model bert-large-uncased 
--output_dir OUTPUT_DIR/rl/
--train_file DATA_DIR/squad-qa/wikipedia-train.json
--max_seq_length MAX_SEQ_LENGTH
--max_query_length MAX_QUERY_LENGTH
--do_train
--do_validate
--do_lower_case
--pretrained_model_path OUTPUT_DIR/pretrained/best_BERT_model.bin
--recur_type RECUR_TYPE
--train_batch_size 8
--learning_rate 1e-5
--num_train_epochs 2.0
--max_read_times MAX_READ_TIMES
--max_answer_length 60

• MAX_SEQ_LENGTH: input sequence length, and is set as 512 in the experiment.

• MAX_QUERY_LENGTH: the maximum length of a query in an input sequence (set as 64).

• RECUR_TYPE: can be "gated" or "lstm" corresponding to Gated and LSTM recurrence respectively.

• MAX_READ_TIMES: the maximum number of segments read from a document. It is set as 3 in the experiment.

Prediction

python3 train/run_RCM_trivia.py 
--bert_model bert-large-uncased 
--output_dir OUTPUT_DIR/rl/
--predict_file DATA_DIR/squad-qa/wikipedia-dev.json
--use_history
--n_history -1
--max_seq_length MAX_SEQ_LENGTH
--max_query_length 64
--doc_stride DOC_STRIDE
--do_predict
--do_lower_case
--recur_type RECUR_TYPE
--predict_batch_size 12
--max_read_times MAX_READ_TIMES
--max_answer_length 60

• Predictions will be saved in OUTPUT_DIR/rl/predictions.json

Evaluation

Download official trivia repo, and go to the repo folder triviqa/. Follow the instruction to evaluate predictions.

python3 -m evaluation.triviaqa_evaluation --dataset_file DATA_DIR/qa/wikipedia-dev.json --prediction_file OUTPUT_DIR/rl/predictions.json

Ablation Study

To evaluate the gain of chunking mechanism, train a model with recurrence but without flexible strides on CoQA dataset:

python3 train/run_RCM_coqa.py 
--bert_model bert-large-uncased 
--output_dir OUTPUT_DIR/pretrained/
--train_file DATA_DIR/coqa.train.json
--use_history
--n_history -1
--max_seq_length MAX_SEQ_LENGTH
--max_query_length MAX_QUERY_LENGTH
--doc_stride DOC_STRIDE
--do_train
--do_validate
--do_lower_case
--recur_type RECUR_TYPE
--supervised_pretraining
--train_batch_size 8
--learning_rate 3e-5
--num_train_epochs 2.0
--max_read_times 3
--max_answer_length 30

If you have any questions, please contact Hongyu Gong (hongyugong6@gmail.com).

If you use our code, please cite our work:

Hongyu Gong, Yelong Shen, Dian Yu, Jianshu Chen and Dong Yu, "Recurrent Chunking Mechanisms for Long-Text Machine Reading Comprehension", in Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) 2020 Jul (pp. 6751--6761).

@inproceedings{gong-etal-2020-recurrent, title = "Recurrent Chunking Mechanisms for Long-Text Machine Reading Comprehension", author = "Gong, Hongyu and Shen, Yelong and Yu, Dian and Chen, Jianshu and Yu, Dong", booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics", month = jul, year = "2020", address = "Online", publisher = "Association for Computational Linguistics", url = "https://www.aclweb.org/anthology/2020.acl-main.603", pages = "6751--6761"}