Rationale Prediction

PyTorch reimplementation of Rationalizing Neural Predictions

Using Gumbel Softmax instead of REINFORCE.

Requirements

1. Pytorch 1.0
2. AllenNLP
3. nltk
4. tqdm

Usage

See models/train.py for details of commandline options.

Dataset

1. Rotten Tomatoes;
2. Congress Dataset (ETH internal only);
3. BeerReview Dataset (TODO).

Result

TODO