EvEntS ReaLM

This repo contains the code for our paper EvEntS ReaLM: Event Reasoning of Entity States via Language Models, to-appear in EMNLP 2022. We provide code to reproduce the experiments described in the paper. We also provide the preprocessed data used in the experiments to facilitate reproducibility of our experiments.

Requirements

We use the PyTorch library in all our experiments. Our zero-prompt experiments are based on the Simple Transforemrs librarry. We adapt the Huggingface Transformers library for single, all, and k attribute prompt experiments.

For the single-Attribute prompt experiments with Roberta, we had to modify the original Transformer library. We therefore provide a full copy of the library with our changes. To reproduce these experiments, it is necessary to install this library. Move to the transformers-single-all-attribute-prompt-experiments directory and then execute pip install -e .

Data, Outputs, and Checkpoints

We provide the preprocessed data for most experiments in the data directory. This data was adapted from the original PiGLET and OpenPI datasets. If you use the data make sure to cite the original work.

In addition, data, outputs, and checkpoint for the k-attribute prompt model available at the following drive link: https://drive.google.com/drive/folders/1TVxD0biBa04OKZTugV6iBZtYrNFPYIaM?usp=sharing

Experiments

The /scripts directory contains the commands to run the experiments described in the paper, separated based on the dataset (PiGLET vs OpenPI).

Citation

@article{spiliopoulou2022events,
  title={EvEntS ReaLM: Event Reasoning of Entity States via Language Models},
  author={Spiliopoulou, Evangelia and Pagnoni, Artidoro and Bisk, Yonatan and Hovy, Eduard},
  journal={arXiv preprint arXiv:2211.05392},
  year={2022}
}