TaSc

Repository for the paper "Improving Attention Interpretability with Task-specific Information in Text Classification", G.Chrysostomou and N.Aletras, to appear at ACL2021. Pre-print available at this link

Prerequisites

Install necessary packages by using the files conda_reqs.txt and pip_reqs.txt

conda create --name tasc --file  conda_reqs.txt
conda activate tasc
pip install -r pip_reqs.txt
python -m spacy download en

Downloading Task Data

You can run the jupyter notebooks found under tasks/task_name/*ipynb to generate a filtered, processed csv file and a pickle file used for trainining the models.

Training and Evaluating the models

You can train and save the models with train_eval_bc.py script with the following options:

• dataset : {sst, twitter, mimicanemia, imdb, agnews}
• encoder : {lstm, gru, mlp, cnn, bert}
• data_dir : directory where task data is
• model_dir : directory for saved models
• experiments_dir : direcrory to save experiment results
• mechanisms: selection of attention mechanism with options {tanh, dot}
• operation : operation over tasc score generation with option {sum-over, max-pool, mean-pool}
• lin : apply Lin-TaSc
• feat : apply Feat-TaSc
• conv : apply Conv-TaSc
• speed_up : In case of low resource scenarios, you can use this option to speed up the fraction of token experiments by searching every 2% of the sequence instead of every token, with results being similar (almost identical) to those of the paper.

Example script (without TaSc):

python train_eval_bc.py -dataset sst 
			-encoder lstm 
			-mechanism dot 
			-data_dir data/ 
			-model_dir models/ 

Example script (with Lin-TaSc):

python train_eval_bc.py -dataset sst 
			-encoder lstm 
			-mechanism dot 
			-data_dir data/ 
			-model_dir models/ 
			-lin

Summarising results

Following the evaluation for multiple models / attention mechanisms , with and without TaSc, you can use produce_reports.py to create tables in latex and as csv for the full stack of results (as a comparison), a table for comparing with other explanation techniques, results across attention mechanism, encoder and dataset.

The script can be run with the following options:

• datasets: list of datasets with permissible datasets listed above
• encoders: list of encoders with permissible encoders listed above
• experiments_dir: directory that contains saved results to use for summarising followed by /
• mechanisms: selection of attention mechanism with options {Tanh, Dot}
• tasc_ver : tasc version with options {lin, feat, conv}

To generate radar plots you can run python produce_graphs.py(produce_graphs.py).