There was a minor issue in the previous version of our paper. Specifically, Equation 4 should read as follows: $\mathop{\min}\limits_{\theta_{G_1},\ldots,\theta_{G_n}, \theta_P}\sum\limits_{(X,Y)\sim \mathcal{D}} H(Y,\hat{Y})=\sum\limits_{(X,Y)\sim \mathcal{D}} \sum\limits_{i=1}^{n} H(Y,f_{P}(f_{G_i}(X)))$.
We have rectified this error in our arxiv version.

We highly recommend for reading a companion piece, DR, to this paper.

MGR: Multi-generator based Rationalization

This repo contains Pytorch implementation of MGR: Multi-generator based Rationalization (ACL 2023 oral).

If the code has any bugs, please open an issue. We will be grateful for your help.

We have provided some tips for you to better understand our code and build your method on top of it: tips. If you're still having trouble reproducing your code after reading these tips, I'd be happy to set up a video conference to help you.

Preparing code can be an exhausting task. We would be grateful if you would star this repo before cloning it. If you have any questions, just open an issue or send us an e-mail.
Thank you!

Environments

RTX3090

Create an environment with: conda create -n MGR python=3.7.13

Then activate it: conda activate MGR

Install pytorch: conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.3 -c pytorch

Install other packages: pip install -r requirements.txt

Due to different versions of torch, you may need to replace "cls_loss = args.cls_lambda * F.cross_entropy(forward_logit, labels)" with "cls_loss = args.cls_lambda * F.cross_entropy(forward_logit, labels.long())"

Datasets

For Beer Reviews, you should first obtain authorization for this dataset from the original author.

Beer Reviews: you can get it here. Then place it in the ./data/beer directory.
Hotel Reviews: you can get it here. Then find hotel_Location.train, hotel_Location.dev, hotel_Service.train, hotel_Service.dev, hotel_Cleanliness.train, hotel_Cleanliness.dev from data/oracle and put them in the ./data/hotel directory. Find hotel_Location.train, hotel_Service.train, hotel_Cleanliness.train from data/target and put them in the ./data/hotel/annotations directory.
Word embedding: glove.6B.100d.txt. Then put it in the ./data/hotel/embeddings directory.

Running Example

Correlated Beer

Appearance:

For sparsity $S\approx 10$:
python -u norm_beer.py --correlated 1 --lr 0.00003 --batch_size 128 --gpu 0 --sparsity_percentage 0.083 --epochs 400 --aspect 0

For sparsity $S\approx 20$:
python -u norm_beer.py --correlated 1 --lr 0.00003 --batch_size 128 --gpu 0 --sparsity_percentage 0.173 --epochs 400 --aspect 0

For sparsity $S\approx 30$:
python -u norm_beer.py --correlated 1 --lr 0.00003 --batch_size 128 --gpu 0 --sparsity_percentage 0.273 --epochs 400 --aspect 0

Hotel

cleanliness:
python -u norm_beer.py --data_type hotel --lr 0.00007 --batch_size 1024 --gpu 0 --sparsity_percentage 0.1 --sparsity_lambda 10 --continuity_lambda 10 --epochs 800 --aspect 2

Notes: "--sparsity_percentage 0.173" means "$s=0.173$" in Eq.3 (But the actual sparsity is different from $s$. See the results below.). "--sparsity_lambda 10 --continuity_lambda 10 " means $\lambda_1=10, \lambda_2=10$. "--epochs 400" means we run 400 epochs and take the results when the "dev_acc" is best.

When you change the random seed, you need to adjust the "sparsity_percentage" according to the actual sparsity on the test set.

Result

You will get a result like "best_dev_epoch=194" at last. Then you need to find the result corresponding to the epoch with number "194".
For Beer-Appearance, you may get a result like:

Train time for epoch #194 : 23.373237 second
traning epoch:194 recall:0.7490 precision:0.8077 f1-score:0.7772 accuracy:0.7853
Validate
dev epoch:194 recall:0.7430 precision:0.7528 f1-score:0.7479 accuracy:0.7495
Validate Sentence
dev dataset : recall:0.6829 precision:0.7212 f1-score:0.7015 accuracy:0.7094
Annotation
annotation dataset : recall:0.8505 precision:0.9987 f1-score:0.9187 accuracy:0.8515
The annotation performance: sparsity: 20.2753, precision: 76.3473, recall: 83.6095, f1: 79.8135
Annotation Sentence
annotation dataset : recall:0.8548 precision:0.9987 f1-score:0.9212 accuracy:0.8558
Rationale
rationale dataset : recall:0.8375 precision:0.9974 f1-score:0.9105 accuracy:0.8376

The last line "The annotation performance: sparsity: 20.2753, precision: 76.3473, recall: 83.6095, f1: 79.8135 " indicates the overlap between the selected tokens and human-annotated rationales.

Citation

@inproceedings{liu-etal-2023-mgr,
title = "{MGR}: Multi-generator Based Rationalization",
author = "Liu, Wei and
Wang, Haozhao and
Wang, Jun and
Li, Ruixuan and
Li, Xinyang and
Zhang, YuanKai and
Qiu, Yang",
booktitle = "Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)",
month = july,
year = "2023",
address = "Toronto, Canada",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.acl-long.715",
pages = "12771--12787",
}

Acknowledgement

The code is largely based on Car and DMR. Most of the hyperparameters (e.g. the '--cls_lambda'=0.9) are also from them. We are grateful for their open source code.