Skip to content

[NAACL 2021] QAGNN: Question Answering using Language Models and Knowledge Graphs πŸ€–

License

Notifications You must be signed in to change notification settings

michiyasunaga/qagnn

Repository files navigation

QA-GNN: Question Answering using Language Models and Knowledge Graphs

This repo provides the source code & data of our paper: QA-GNN: Reasoning with Language Models and Knowledge Graphs for Question Answering (NAACL 2021).

@InProceedings{yasunaga2021qagnn,
  author =  {Michihiro Yasunaga and Hongyu Ren and Antoine Bosselut and Percy Liang and Jure Leskovec},
  title =   {QA-GNN: Reasoning with Language Models and Knowledge Graphs for Question Answering},
  year =    {2021},  
  booktitle = {North American Chapter of the Association for Computational Linguistics (NAACL)},  
}

Webpage: https://snap.stanford.edu/qagnn

Usage

0. Dependencies

Run the following commands to create a conda environment (assuming CUDA10.1):

conda create -n qagnn python=3.7
source activate qagnn
pip install torch==1.8.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html
pip install transformers==3.4.0
pip install nltk spacy==2.1.6
python -m spacy download en

# for torch-geometric
pip install torch-scatter==2.0.7 -f https://pytorch-geometric.com/whl/torch-1.8.0+cu101.html
pip install torch-sparse==0.6.9 -f https://pytorch-geometric.com/whl/torch-1.8.0+cu101.html
pip install torch-geometric==1.7.0 -f https://pytorch-geometric.com/whl/torch-1.8.0+cu101.html

1. Download data

We use the question answering datasets (CommonsenseQA, OpenBookQA) and the ConceptNet knowledge graph. Download all the raw data by

./download_raw_data.sh

Preprocess the raw data by running

python preprocess.py -p <num_processes>

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 by

./download_preprocessed_data.sh

πŸ”΄ NEWS (Add MedQA-USMLE). Besides the commonsense QA datasets (CommonsenseQA, OpenBookQA) with the ConceptNet knowledge graph, we added a biomedical QA dataset (MedQA-USMLE) with a biomedical knowledge graph based on Disease Database and DrugBank. You can download all the data for this from [here]. Unzip it and put the medqa_usmle and ddb folders inside the data/ directory. While this data is already preprocessed, we also provide the preprocessing scripts we used in utils_biomed/.

The resulting file structure will look like:

.
β”œβ”€β”€ README.md
β”œβ”€β”€ data/
    β”œβ”€β”€ cpnet/                 (prerocessed ConceptNet)
    β”œβ”€β”€ csqa/
        β”œβ”€β”€ train_rand_split.jsonl
        β”œβ”€β”€ dev_rand_split.jsonl
        β”œβ”€β”€ test_rand_split_no_answers.jsonl
        β”œβ”€β”€ statement/             (converted statements)
        β”œβ”€β”€ grounded/              (grounded entities)
        β”œβ”€β”€ graphs/                (extracted subgraphs)
        β”œβ”€β”€ ...
    β”œβ”€β”€ obqa/
    β”œβ”€β”€ medqa_usmle/
    └── ddb/

2. Train QA-GNN

For CommonsenseQA, run

./run_qagnn__csqa.sh

For OpenBookQA, run

./run_qagnn__obqa.sh

For MedQA-USMLE, run

./run_qagnn__medqa_usmle.sh

As configured in these scripts, the model needs two types of input files

  • --{train,dev,test}_statements: preprocessed question statements in jsonl format. This is mainly loaded by load_input_tensors function in utils/data_utils.py.
  • --{train,dev,test}_adj: information of the KG subgraph extracted for each question. This is mainly loaded by load_sparse_adj_data_with_contextnode function in utils/data_utils.py.

Note: We find that training for OpenBookQA is unstable (e.g. best dev accuracy varies when using different seeds, different versions of the transformers / torch-geometric libraries, etc.), likely because the dataset is small. We suggest trying out different seeds. Another potential way to stabilize training is to initialize the model with one of the successful checkpoints provided below, e.g. by adding an argument --load_model_path obqa_model.pt.

3. Evaluate trained model

For CommonsenseQA, run

./eval_qagnn__csqa.sh

Similarly, for other datasets (OpenBookQA, MedQA-USMLE), run ./eval_qagnn__obqa.sh and ./eval_qagnn__medqa_usmle.sh. You can download trained model checkpoints in the next section.

Trained model examples

CommonsenseQA

Trained model In-house Dev acc. In-house Test acc.
RoBERTa-large + QA-GNN [link] 0.7707 0.7405

OpenBookQA

Trained model Dev acc. Test acc.
RoBERTa-large + QA-GNN [link] 0.6960 0.6900

MedQA-USMLE

Trained model Dev acc. Test acc.
SapBERT-base + QA-GNN [link] 0.3789 0.3810

Note: The models were trained and tested with HuggingFace transformers==3.4.0.

Use your own dataset

  • Convert your dataset to {train,dev,test}.statement.jsonl in .jsonl format (see data/csqa/statement/train.statement.jsonl)
  • Create a directory in data/{yourdataset}/ to store the .jsonl files
  • Modify preprocess.py and perform subgraph extraction for your data
  • Modify utils/parser_utils.py to support your own dataset

Acknowledgment

This repo is built upon the following work:

Scalable Multi-Hop Relational Reasoning for Knowledge-Aware Question Answering. Yanlin Feng*, Xinyue Chen*, Bill Yuchen Lin, Peifeng Wang, Jun Yan and Xiang Ren. EMNLP 2020.
https://github.com/INK-USC/MHGRN

Many thanks to the authors and developers!