BERT-Analysis-RCQA

Official code for paper Towards Interpreting BERT for Reading Comprehension Based QA.

This code was developed on top of the official BERT code released by Google.

Experimental Setup

The base model checkpoints for BERT can be downloaded from the official git repository. We used uncased_L-12_H-768_A-12, which has the following configuration: 12-layer, 768-hidden, 12-heads, 110M parameters. All codes have to be run inside the bert directory.

We used the SQuAD-V1 and Self-RC (DuoRC) datasets - all data processing is taken care of by the code itself, before training or evaluation.

We used tensorflow-v1.14 with Python 2 for training/evaluation, and Python 3 for all analyses.

To finetune the BERT model on SQuAD/DuoRC:

Use run_squad_infer.py for SQuAD and duorc_infer.py for DuoRC.

export BERT_BASE_DIR=/path/to/bert/uncased_L-12_H-768_A-12
export SQUAD_DIR=/path/to/json-datasets

python -u run_squad_infer.py 
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --do_train=False \
  --train_file=$SQUAD_DIR/train-json-file \
  --do_predict=True \
  --predict_file=$SQUAD_DIR/dev-json-file \
  --train_batch_size=6 \
  --learning_rate=3e-5 \
  --num_train_epochs=2.0 \
  --max_seq_length=384 \
  --doc_stride=128 \
  --output_dir=/path/to/new/model/folder

To evaluate the BERT model's predictions:

Use evaluate-v2.0.py for SQuAD and duorc_evaluate-v2.0.py for DuoRC.

python -u evaluate-v2.0.py /path/to/dataset/json /path/to/predictions/json --checkpoint_dir=/path/to/checkpoints/folder

To generate integrated gradient scores for each layer:

• set layer number from 0 to 11 in do_integrated_grad
• predict_batch_size must always be set to 1 in this code
• The IG scores are stored in a folder called ig_scores (for SQuAD) and ig_scores_duorc (for DuoRC), in files named in the format importance_scores_ig_layer_number.npy
• Embeddings for the first 200 datapoints are stored in a folder called embs (for SQuAD) and embs_duorc (for DuoRC), in files named in the format emb_enclayer_layer_number.npy
• The tokenized 'QN [SEP] PASSAGE' will be stored in output_dir/qn_and_doc_tokens.npy
• Use bert_dec_flips.py for SQuAD and duorc_dec_flips.py for DuoRC (since the datasets need to be processed differently).

export BERT_BASE_DIR=/path/to/bert/uncased_L-12_H-768_A-12
export SQUAD_DIR=/path/to/json-datasets

python -u bert_dec_flips.py \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --do_train=False \
  --train_file=$SQUAD_DIR/train-json-file \
  --do_predict=True \
  --predict_file=$SQUAD_DIR/dev-json-file \
  --train_batch_size=6 \
  --learning_rate=3e-5 \
  --num_train_epochs=2.0 \
  --max_seq_length=384 \
  --doc_stride=128 \
  --predict_batch_size=1 \
  --do_integrated_grad=11 \
  --output_dir=/path/to/new/model/folder

Codes to generate Jensen-Shannon graphs and save them:

Use jensen_shannon.py and graph_js.py.

Code to generate t-SNE plots:

Use tsne.py (this code uses the embeddings saved by the IG scores code above).

Code to analyze quantifier questions:

Use quantifier_questions_analysis.py.

Contact

For any questions, please contact sahanjich@gmail.com.