thcrwi / interpret_bert Goto Github PK

Code used in our ACL'19 paper for interpreting BERT model.

• PyTorch
• pytorch-pretrained-BERT
• SentEval
• spaCy (for dependency tree visualization)

• Navigate:

cd chunking/

• Download the train set from CoNLL-2000 chunking corpus:

wget https://www.clips.uantwerpen.be/conll2000/chunking/train.txt.gz
gunzip train.txt.gz

The last command replaces train.txt.gz file with train.txt file.

• Extract BERT features for chunking related tasks (clustering and visualization):

python extract_features.py --train_file train.txt --output_file chunking_rep.json

• Run t-SNE of span embeddings for each BERT layer (Figure 1):

python visualize.py --feat_file chunking_rep.json --output_file_prefix tsne_layer_

This would create one t-SNE plot for each BERT layer and stores as pdf (e.g. tsne_layer_0.pdf).

• Run KMeans to evaluate the clustering performance of span embeddings for each BERT layer (Table 1):

python cluster.py --feat_file chunking_rep.json

• Navigate:

cd probing/

• Download the data files for 10 probing tasks (e.g. tree_depth.txt)
• Extract BERT features for sentence level probing tasks (tree_depth in this case):

python extract_features.py --data_file tree_depth.txt --output_file tree_depth_rep.json

In the above command, append --untrained_bert flag to extract untrained BERT features.

• Train the probing classifier for a given BERT layer (indexed from 0) and evaluate the performance (Table 2):

python classifier.py --labels_file tree_depth.txt --feats_file tree_depth_rep.json --layer 0

We use the hyperparameter search space recommended by SentEval.

• Navigate:

cd sva/

• Download the data file for SVA task and extract it.
• Extract BERT features for SVA task:

python extract_features.py --data_file agr_50_mostcommon_10K.tsv --output_folder ./

• Train the binary classifier for a given BERT layer (indexed from 0) and evaluate the performance (Table 3):

python classifier.py --input_folder ./ --layer 0

We use the hyperparameter search space recommended by SentEval.

• Navigate:

cd tpdn/

• Download the SNLI 1.0 corpus and extract it.
• Extract BERT features for premise sentences present in SNLI:

python extract_features.py --input_folder . --output_folder .

• Train the Tensor Product Decomposition Network (TPDN) to approximate a given BERT layer (indexed from 0) and evaluate the performance (Table 4):

python approx.py --input_folder . --output_folder . --layer 0

Check --role_scheme and --rand_tree flags for setting the role scheme.

• Induce dependency parse tree from attention weights for a given attention head and BERT layer (both indexed from 1) (Figure 2):

python induce_dep_trees.py --sentence text "The keys to the cabinet are on the table" --head_id 11 --layer_id 2 --sentence_root 6 

This repository would not be possible without the efforts of the creators/maintainers of the following libraries:

• pytorch-pretrained-BERT from huggingface
• SentEval from facebookresearch
• bert-syntax from yoavg
• tpdn from tommccoy1
• rnn_agreement from TalLinzen
• Chu-Liu-Edmonds from bastings

This repository is GPL-licensed.