Skip to content

Latest commit

 

History

History
115 lines (92 loc) · 5.79 KB

README.md

File metadata and controls

115 lines (92 loc) · 5.79 KB
Raw

Faithful Knowledge Graph Explanations in Commonsense Question Answering

This repo accompanies our EMNLP 2022 paper Faithful Knowledge Graph Explanations in Commonsense Question Answering. Here, we provide the modifications to MHGRN and QA-GNN that we used to produce the results in the paper to increase the faithfulness of the explanations.

If you use this code in your work, please cite our paper:

@inproceedings{aglionby-teufel-2022-faithful,
    title = "Faithful Knowledge Graph Explanations in Commonsense Question Answering",
    author = "Aglionby, Guy and Teufel, Simone",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-main.743",
    pages = "10811--10817",
}

Thanks to the authors of MHGRN and QA-GNN, whose code we built on in this repository.

Usage

0. Dependencies

Run the following commands to create a conda environment (assumes CUDA10; requires gcc 7.x or lower):

conda create -n krqa-38 python=3.8 numpy matplotlib ipython
conda activate krqa-38
conda install pytorch cudatoolkit=10.1 -c pytorch
pip install spacy==2.1.6 nltk==3.5 tqdm transformers==4.2.2 networkx==2.6.3 jsonlines pandas==1.2.2 datasets==1.3.0 numpy matplotlib tensorboard
python -m spacy download en_core_web_sm
pip install torch-sparse==0.6.1 -f https://pytorch-geometric.com/whl/torch-1.8.0+cu101.html
pip install torch-scatter==2.0.4 -f https://pytorch-geometric.com/whl/torch-1.8.0+cu101.html
pip install torch-geometric==1.6.0 -f https://pytorch-geometric.com/whl/torch-1.8.0+cu101.html

Cuda 11.1:

conda create -n 38-cu111-new python=3.8 numpy matplotlib ipython
conda activate 38-cu111-new
conda install pytorch=1.8.1 cudatoolkit=11.1 pytorch-sparse -c pytorch -c conda-forge -c pyg
pip install spacy==2.3.7 nltk==3.5 tqdm transformers==4.2.2 networkx==2.6.3 jsonlines pandas==1.2.2 datasets==1.3.0 numpy matplotlib tensorboard
python -m spacy download en_core_web_sm
pip install torch-geometric==1.6.0 --no-cache-dir -f https://pytorch-geometric.com/whl/torch-1.8.1+cu111.html

1. Download data

This step exactly follows QA-GNN. We use the question answering datasets (CommonsenseQA, OpenBookQA) and the ConceptNet knowledge graph. Download all the raw data by

./download_raw_data.sh

The data needs to be pre-processed before using it with the model. One aspect of this is that QA-GNN requires KG concepts identified by the subgraph finding process (sometimes known as schema graph construction) to be rated for relevance to the question/answer pair. This can be done in two ways: using the --use-lm-scoring flag with preprocess.py, or running pre-processing without LM scoring and then using the add_lm_score_to_graph.py script to post-hoc add the LM scores.

# Preprocess ConceptNet
python preprocess.py -p <num_processes> --run graph
# Preprocess the two datasets (using both ways of adding LM scoring)
python preprocess.py -p <num_processes> --run csqa --use-lm-scoring
python preprocess.py -p <num_processes> --run obqa
python add_lm_score_to_graph.py data/obqa-cpnet

The script will:

  • Setup ConceptNet (e.g., extract English relations from ConceptNet, merge the original 42 relation types into 17 types)
  • Convert the QA datasets into .jsonl files (e.g., stored in data/csqa/statement/)
  • Identify all mentioned concepts in the questions and answers
  • Extract subgraphs for each q-a pair

TL;DR (Skip above steps and just get preprocessed data). The preprocessing may take long. For your convenience, you can download all the processed data (provided by QA-GNN authors) by

./download_preprocessed_data.sh

2. Training & ablations

The modifications proposed in our paper are activated by passing flag options to the relevant scripts (grn.py for MHGRN; qagnn.py for QA-GNN). The options are given to the --ablation flag separated by spaces.

Model Flag Description
MHGRN detach_s_all Freeze the text encoder
MHGRN no_s_in_final_mlp Don't use the sentence embedding in the final MLP
QA-GNN detach_s_all Freeze the text encoder
QA-GNN no_s_in_final_mlp_direct Don't use the sentence embedding directly from the text encoder in the final MLP
QA-GNN no_s_in_final_mlp_from_graph Don't use the sentence embedding that was used in the GNN in the final MLP

In the paper, the -Embed experiments correspond to the no_s_in_final_mlp* flags. -Train TE. additionally uses the detach_s_all flag.

scripts/original contains the scripts to train the models for OpenBookQA. Scripts for CommonsenseQA are easily created by substituting csqa for obqa.

3. Swapping KGs between questions

In the paper, we swap the KGs between questions to quantify the possible faithfulness of explanations. Use the generated_swapped_kgs.sh script to generate new datasets with swapped KGs.

You can then use the scripts in scripts/swapped to train and evaluate the models on the swapped datasets.