Official implementation of NeurIPS'2022 paper "DARE:Disentanglement-Augmented Rationale Extraction".
We provide toy simulations in ./Simulated Studies/club_nce.ipynb to show the estimation performance of CLUB_NCE and other MI estimators.
To train DARE on a single aspect, e.g. aspect 0 (look):
python -m latent_rationale.beer.dare \
--model latent \
--aspect 0 \
--epochs 50 \
--lr 0.00012 \
--upper_bound 0.01 \
--batch_size 200 \
--train_path ./beer/reviews.aspect0.train.txt.gz \
--dev_path ./beer/reviews.aspect0.heldout.txt.gz \
--test_path ./beer/annotations.json \
--scheduler exponential \
--save_path ./dare_a0 \
--dependent-z \
--selection 0.13 --lasso 0.02
The backbone of our code is referenced from codes released by HardKuma, CLUB and SMILE. Thank you for their sharing !
@inproceedings{yuedare,
title={DARE: Disentanglement-Augmented Rationale Extraction},
author={Yue, Linan and Liu, Qi and Du, Yichao and An, Yanqing and Wang, Li and Chen, Enhong},
booktitle={Advances in Neural Information Processing Systems},
year={2022}
}