Introduction

This is the project repo for our Findings of EMNLP'22 paper: https://arxiv.org/abs/2210.12902. We borrow a large part of codes from the original ESTER dataset repository: https://github.com/PlusLabNLP/ESTER. Pls refer to the official dataset page for dataset details. Only the training and development set are used in our paper.

Models

I. Install packages.

We list the packages in our environment in env.yml file for your reference.

II. Train and test

1. Fine-tuned models.

We provide several fine-tuned models for quick usage.

Models: https://drive.google.com/drive/folders/1ljx4pgpy0ocyHN6pGBjLTuHaG7Dd0RTu?usp=sharing
Extractive QA: roberta-large_IO_prefix_transE. Download it to ./output/spanqa/.
Generative QA: unifiedqa-t5-large_prefix_transE. Download it to ./output/allenai/.

2. Train from scratch

Run export PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python.
Run bash ./code/run_span_pred.sh and bash ./code/run_ans_generation.sh.

3. Test on dev set

Run bash ./code/eval_span_pred.sh and bash ./code/eval_ans_generation.sh.