This repo provides the source code & data of our paper "DRAGON: Deep Bidirectional Language-Knowledge Graph Pretraining" (NeurIPS 2022).

DRAGON is a new foundation model (improvement of BERT) that is pre-trained jointly from text and knowledge graphs for improved language, knowledge and reasoning capabilities. Specifically, it was trained with two simultaneous self-supervised objectives, language modeling and link prediction, that encourage deep bidirectional reasoning over text and knowledge graphs.

DRAGON can be used as a drop-in replacement for BERT. It achieves better performance in various NLP tasks, and is particularly effective for knowledge and reasoning-intensive tasks such as multi-step reasoning and low-resource QA.

Run the following commands to create a conda environment:

conda create -y -n dragon python=3.8
conda activate dragon
pip install torch==1.10.1+cu113 torchvision -f https://download.pytorch.org/whl/cu113/torch_stable.html
pip install transformers==4.9.1 wandb nltk spacy==2.1.6
python -m spacy download en
pip install scispacy==0.3.0
pip install https://s3-us-west-2.amazonaws.com/ai2-s2-scispacy/releases/v0.3.0/en_core_sci_sm-0.3.0.tar.gz
pip install torch-scatter==2.0.9 torch-sparse==0.6.12 torch-geometric==2.0.0 -f https://pytorch-geometric.com/whl/torch-1.10.1+cu113.html

You can download pretrained DRAGON models below. Place the downloaded model files under ./models

You can download all the preprocessed data from [here]. This includes the ConceptNet knowledge graph as well as CommonsenseQA, OpenBookQA and RiddleSense datasets. Specifically, run:

wget https://nlp.stanford.edu/projects/myasu/DRAGON/data_preprocessed.zip
unzip data_preprocessed.zip
mv data_preprocessed data

(Optional) If you would like to preprocess the raw data from scratch, you can download the raw data – ConceptNet Knowledge graph, CommonsenseQA, OpenBookQA – by:

./download_raw_data.sh

To preprocess the raw data, run:

CUDA_VISIBLE_DEVICES=0 python preprocess.py -p <num_processes> --run common csqa obqa

You can specify the GPU you want to use in the beginning of the command CUDA_VISIBLE_DEVICES=.... 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

You can download all the preprocessed data from [here]. This includes the UMLS biomedical knowledge graph and MedQA dataset.

(Optional) If you would like to preprocess MedQA from scratch, follow utils_biomed/preprocess_medqa.ipynb and then run

CUDA_VISIBLE_DEVICES=0 python preprocess.py -p <num_processes> --run medqa

.
├── README.md
├── models/
    ├── general_model.pt
    ├── biomed_model.pt

└── data/
    ├── cpnet/                 (preprocessed ConceptNet KG)
    └── csqa/
        ├── train_rand_split.jsonl
        ├── dev_rand_split.jsonl
        ├── test_rand_split_no_answers.jsonl
        ├── statement/             (converted statements)
        ├── grounded/              (grounded entities)
        ├── graphs/                (extracted subgraphs)
        ├── ...
    ├── obqa/
    ├── umls/                  (preprocessed UMLS KG)
    └── medqa/

To train DRAGON on CommonsenseQA, OpenBookQA, RiddleSense, MedQA, run:

scripts/run_train__csqa.sh
scripts/run_train__obqa.sh
scripts/run_train__riddle.sh
scripts/run_train__medqa.sh

(Optional) If you would like to pretrain DRAGON (i.e. self-supervised pretraining), run

scripts/run_pretrain.sh

As a quick demo, this script uses sentences from CommonsenseQA as training data. If you wish to use a larger, general corpus like BookCorpus, follow Section 5 (Use your own dataset) to prepare the training data.

For CommonsenseQA, OpenBookQA, RiddleSense, MedQA, run:

scripts/run_eval__csqa.sh
scripts/run_eval__obqa.sh
scripts/run_eval__riddle.sh
scripts/run_eval__medqa.sh

You can download trained model checkpoints in the next section.

CommonsenseQA

OpenBookQA

RiddleSense

MedQA

Note: The models were trained and tested with HuggingFace transformers==4.9.1.

• 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

If you find our work helpful, please cite the following:

@InProceedings{yasunaga2022dragon,
  author =  {Michihiro Yasunaga and Antoine Bosselut and Hongyu Ren and Xikun Zhang and Christopher D. Manning and Percy Liang and Jure Leskovec},
  title =   {Deep Bidirectional Language-Knowledge Graph Pretraining},
  year =    {2022},  
  booktitle = {Neural Information Processing Systems (NeurIPS)},  
}

This repo is built upon the following works:

GreaseLM: Graph REASoning Enhanced Language Models for Question Answering
https://github.com/snap-stanford/GreaseLM

QA-GNN: Question Answering using Language Models and Knowledge Graphs
https://github.com/michiyasunaga/qagnn