Incorporating-Priors

This repository contains implementation of paper -- Incorporating Priors with Feature Attribution on Text Classification (ACL2019)

How to train a model

python train.py --num_classes=2 --output_dir='/tmp/attr/' --target_words_file='./data/wiki/identity.txt'

By default, the command trains the model with joint loss (attribution + classification) on ./data/wiki/wiki_train.txt.

Check

python train.py -h

for other settings.

How to predict and generate global attributions (average of local attributions).

python explain.py --model_dir=/tmp/attr --pred_output=/tmp/pred.txt

Check

python explain.py -h

for other settings.