README.md

CGR

Code for our EMNLP Findings 2021 paper,

Exploiting Reasoning Chains for Multi-hop Science Question Answering

Weiwen Xu, Yang Deng, Huihui Zhang, Deng Cai and Wai Lam.

Data Preparation

We present the results on OpenBookQA and ARC-Challenge in our paper. Due to the license issue, please directly download the datasets from their corresponding websites.

Data Annotation

We use this repo as our hypothesis generator and AMR-gs as our AMR parser. Please follow their instructions to annotate hypothesis and AMR for the datasets respectively.

Once annotated, please organize the annotated files in the following directory (e.g. OpenBookQA)

- Data/
    - OpenBook/
        - train-complete.jsonl (train/dev/test original datasets)
        - dev-complete.jsonl
        - test-complete.jsonl
        - openbook.txt
        - ARC_Corpus.txt
        - train-hypo.txt (train/dev/test hypotheses)
        - dev-hypo.txt
        - test-hypo.txt
        - train-amr.txt (train/dev/test AMRs)
        - dev-amr.txt
        - test-amr.txt
        - core-amr.txt (core fact AMRs from open-book)
        - comm-amr.txt (common fact AMRs from ARC-Corpus)

Please use scripts/clean_corpus.py to clean the ARC-Corpus to remove noisy sentences.

Preprocessing

1. Add hypothesis to the original datasets:

bash enhance_hypo.sh

2. Add AMR to the hypothesis-enhanced datasets as well as cache all facts AMR:

bash enhance_AMR.sh

3. Cache all dense vectors for evidence facts:

bash cache_vector.sh

Once get all preprocessed, you will get the following directory:

- Data/
    - begin/
        - obqa/
            - train.jsonl
            - dev.jsonl
            - test.jsonl
            - core.dict (AMR file for core facts)
            - core.npy (vector file for core facts)
            - obqa.dict
            - obqa.npy

Training

bash finetune.sh

Citation

If you find this work useful, please star this repo and cite our paper as follows:

@article{xu2021exploiting,
  title={Exploiting Reasoning Chains for Multi-hop Science Question Answering},
  author={Xu, Weiwen and Deng, Yang and Zhang, Huihui and Cai, Deng and Lam, Wai},
  journal={arXiv preprint arXiv:2109.02905},
  year={2021}
}