ExEnt

This repository contains the official code for models described in the paper: "CLUES: A Benchmark for Learning Classifiers using Natural Language Explanations".

CLUES is a benchmark of structured-data classification tasks paired with natural language explanations. In this repository, we provide code for ExEnt (Explanation Entailment), a model capable of zero-shot generalization to novel tasks purely from language explanations.

Model

Setup

Setup environment by running source bin/init.sh. This will

• Download the datasets from: https://clues-benchmark.github.io
• Install and setup environment with correct dependencies.

Training

To train the model, run the following commands:

bash bin/setup.sh
bash bin/train.sh {model_name} {benchmark_type}

Here, {model_name} is one of 'ExEnt', 'RobertaNoExp, 'RobertaExpConcat'. {benchmark_type} refers to one of 'real' or 'syn' to train on CLUES-Real and CLUES-Syn respectively.

The output will be in the experiment directory exp_out/{model_name}/{benchmark_type}/{timestamp}/. Once the model has been trained, the following files can be found in the directory:

exp_out/{model_name}/{benchmark_type}/{timestamp}/
    |
    |__ best_model.pt
    |__ dev_scores.json
    |__ config.json
    |__ src

Note: The best model saved here is based on the averaged validation accuracy over the train tasks.

Zero-Shot Evaluation

To evaluate the model on the test tasks, run the following command:

bash bin/setup.sh
bash bin/test.sh exp_out/{model_name}/{benchmark_type}/{timestamp}/

The test scores for each task will be written in a dictionary under exp_out/{model_name}/{benchmark_type}/{timestamp}/test_scores.json.

Contact

For any doubts or questions regarding the work, please contact Rakesh ([email protected]) or Sayan ([email protected]). For any bug or issues with the code, feel free to open a GitHub issue or pull request.

Citation

@inproceedings{menon2022clues,
    Author = {Menon, Rakesh R. and Ghosh, Sayan and Srivastava, Shashank},
    Title = {{CLUES}: {A} {B}enchmark for {L}earning {C}lassifiers using {N}atural {L}anguage {E}xplanations},
    Year = 2022,
    Journal = {In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (to appear)},
}