PyTorch code for our ACL 2022 paper:
Swarnadeep Saha, Prateek Yadav, and Mohit Bansal
This repository is tested on Python 3.8.3.
You should install the repository on a virtual environment. All dependencies can be installed as follows:
pip install -r requirements.txt
ExplaGraphs dataset can be found inside the data folder. For more details, check out the ExplaGraphs website hosted here.
It contains the training data in train.tsv and the validation samples in dev.tsv.
Each training sample contains four tab-separated entries -- belief, argument, stance label and the explanation graph.
The graph is organized as a bracketed string (edge_1)(edge_2)...(edge_n), where each edge is of the form concept_1; relation; concept_2.
All the negatively perturbed graphs for training our contrastive models can be found in contrastive_data/train.neg_target.
The corresponding gold samples (belief, argument, stance) and the gold graphs are contained in contrastive_data/train.source and contrastive_data/train.target. The files are created in a way to be directly used for training the models.
The positively perturbed graphs can be found in contrastive_data/train.pos_target.
We experiment with rationalizing models that first predict the stance and then conditions on it to generate the graph.
For training the stance prediction model, run the following script
bash scripts/train_stance.sh
The validation samples in contrastive_data/val.source are already appended with the predicted stances from our best model. If you train your own stance prediction model, replace the stances in contrastive_data/val.source with your predictions so that you condition on the predicted stances before generating the graph.
Next, the Max-margin Graph Generation model and the Contrastive model can be trained using the following scripts.
bash scripts/train_graph_max_margin.sh
bash scripts/train_graph_contrastive.sh
All trained models will be saved in the models folder. The scripts to evaluate your models can also be found in the scripts folder.
We'll share our trained models soon. Stay tuned!
You can find the predicted stances and the generated graphs from our max-margin model in output/preds.tsv.
For running the evaluation metrics, refer to the detailed steps outlined in the original ExplaGraphs repository. If you wish to reproduce our results on the validation set of ExplaGraphs, run the evaluation scripts on output/preds.tsv.
Note that the test set is hidden and if you wish to evaluate your own model on it, refer to the instructions here.
@inproceedings{saha2022explagraphsgen,
title={Explanation Graph Generation via Pre-trained Language Models: An Empirical Study with Contrastive Learning},
author={Saha, Swarnadeep and Yadav, Prateek and Bansal, Mohit},
booktitle={ACL},
year={2022}
}
@inproceedings{saha2021explagraphs,
title={ExplaGraphs: An Explanation Graph Generation Task for Structured Commonsense Reasoning},
author={Saha, Swarnadeep and Yadav, Prateek and Bauer, Lisa and Bansal, Mohit},
booktitle={EMNLP},
year={2021}
}