TAT-QA: A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance

TAT-QA (Tabular And Textual dataset for Question Answering) contains 16,552 questions associated with 2,757 hybrid contexts from real-world financial reports.

You can download our TAT-QA dataset via TAT-QA dataset.

For more information, please refer to our TAT-QA website or read our ACL2021 paper PDF.

TagOp Model

Requirements

To create an environment with MiniConda and activate it.

conda create -n tat-qa python==3.7
conda activate tat-qa
pip install -r requirement.txt
pip install torch-scatter -f https://pytorch-geometric.com/whl/torch-1.7.0+${CUDA}.html

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

cd dataset_tagop
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 & Testing

Preprocessing dataset

We heuristicly generate the "facts" and "mapping" fields based on raw dataset, which are stored under the folder of dataset_tagop.

Prepare dataset

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

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

Train & Evaluation

CUDA_VISIBLE_DEVICES=2 PYTHONPATH=$PYTHONPATH:$(pwd) python tag_op/trainer.py --data_dir tag_op/cache/ \
--save_dir ./checkpoint --batch_size 48 --eval_batch_size 8 --max_epoch 50 --warmup 0.06 --optimizer adam --learning_rate 5e-4 \
--weight_decay 5e-5 --seed 123 --gradient_accumulation_steps 4 --bert_learning_rate 1.5e-5 --bert_weight_decay 0.01 \
--log_per_updates 50 --eps 1e-6 --encoder roberta

Testing

CUDA_VISIBLE_DEVICES=2 PYTHONPATH=$PYTHONPATH:$(pwd) python tag_op/predictor.py --data_dir tag_op/cache/ --test_data_dir tag_op/cache/ \\
--save_dir tag_op/ --eval_batch_size 32 --model_path ./checkpoint --encoder roberta

Note: The training process may take around 2 days using a single 32GB v100.

Citation

Please kindly cite our work if you use our dataset or codes, thank you.

@inproceedings{zhu-etal-2021-tat,
    title = "{TAT}-{QA}: A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance",
    author = "Zhu, Fengbin  and
      Lei, Wenqiang  and
      Huang, Youcheng  and
      Wang, Chao  and
      Zhang, Shuo  and
      Lv, Jiancheng  and
      Feng, Fuli  and
      Chua, Tat-Seng",
    booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
    month = aug,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.acl-long.254",
    doi = "10.18653/v1/2021.acl-long.254",
    pages = "3277--3287"
}

Any Question?

For any issues please create an issue here or kindly email us at: Youcheng Huang 1361881994@qq.com or Fengbin Zhu zhfengbin@gmail.com, thank you.

22842219 / TAT-QA