MVGE

Multi-View Graph Representation Learning for Answering Hybrid Numerical Reasoning Question

Please kindly cite our work if you find our codes useful, thank you.

@misc{wei2023multiview,
      title={Multi-View Graph Representation Learning for Answering Hybrid Numerical Reasoning Question}, 
      author={Yifan Wei and Fangyu Lei and Yuanzhe Zhang and Jun Zhao and Kang Liu},
      year={2023},
      eprint={2305.03458},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

requirements

To create an environment with conda and activate it.

conda create -n reghnt python==3.7
conda activate reghnt
pip install torch==1.7.1+cu110 torchvision==0.8.2+cu110 torchaudio==0.7.2 -f https://download.pytorch.org/whl/torch_stable.html     # Adjust according to your CUDA version
pip install allennlp==0.8.4 transformers==4.21.1 nltk==3.5 pandas==1.1.5 numpy==1.21.6
conda install -c dglteam dgl-cuda11.1==0.6.1    # Adjust according your CUDA version
pip install sentencepiece

Next, you should install torch-scatter==2.0.5 (python3.7 CUDA11.1) by Download torch-scatter wheel. (already existed)

Or download other version according to your Python, Pytorch and CUDA version. Then move it to RegHNT/.

pip install torch_scatter-2.0.5-cp37-cp37m-linux_x86_64.whl

We adopt RoBERTa as our encoder to develop our MVGE and use the following commands to prepare RoBERTa model

cd dataset_reghnt
mkdir roberta.large && cd roberta.large
wget -O pytorch_model.bin https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-pytorch_model.bin
wget -O config.json https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-config.json
wget -O vocab.json https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-vocab.json
wget -O merges.txt https://s3.amazonaws.com/models.huggingface.co/bert/roberta-large-merges.txt

Training

Preprocessing dataset

We use the preprocessed data by TagOp Model and they are already in this repository.

Prepare dataset

PYTHONPATH=$PYTHONPATH:$(pwd):$(pwd)/MVGE python mvge/prepare_dataset.py --mode [train/dev/test]

Note: The result will be written into the folder ./cache default.

Train

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=$PYTHONPATH:$(pwd) python mvge/trainer.py --data_dir ./cache/ \
--save_dir ./try --batch_size 48 --eval_batch_size 1 --max_epoch 100 --warmup 0.06 --optimizer adam --learning_rate 1e-4 \
--weight_decay 5e-5 --seed 42 --gradient_accumulation_steps 12 --bert_learning_rate 1e-5 --bert_weight_decay 0.01 \
--log_per_updates 50 --eps 1e-6 --encoder roberta_large --roberta_model dataset_reghnt/roberta.large

Evaluation

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=$PYTHONPATH:$(pwd) python mvge/predictor.py --data_dir ./cache/ \
--test_data_dir ./cache/ --save_dir mvge --eval_batch_size 1 --model_path ./try \
--encoder roberta_large --roberta_model dataset_reghnt/roberta.large --mode dev

python tatqa_eval.py --gold_path=dataset_reghnt/tatqa_dataset_dev.json --pred_path=mvge/pred_result_on_dev.json

Testing

CUDA_VISIBLE_DEVICES=0 PYTHONPATH=$PYTHONPATH:$(pwd) python mvge/predictor.py \
--data_dir ./cache/ --test_data_dir ./cache/ --save_dir mvge \
--eval_batch_size 1 --model_path ./try --encoder roberta_large --roberta_model dataset_reghnt/roberta.large --mode test

Note: The training process may take around 3 days using a single 24GB RTX3090.