next-tat/TAT-DQA

TAT-DQA

• —**Project Page**
• —**Paper - MM 22**
• —**Paper - Arxiv**
• —**Github**
• —**Leaderboard**

TAT-DQA is a large-scale Document VQA dataset, which is constructed by extending the TAT-QA. It aims to stimulate the progress of QA research over more complex and realistic visually-rich documents with rich tabular and textual content, especially those requiring numerical reasoning.

The unique features of TAT-DQA include:

• —The documents in TAT-DQA dataset are sampled from real-world high-quality financial reports and each document contains both tabular and textual data;
• —The average number of words of each document in TAT-DQA is around 550, which is significantly larger than all existing Document VQA datasets.
• —Around 85% of the documents in the dataset have only one page while 15% has multiple pages.
• —Similar to TAT-QA, the answer forms are diverse, including single span, multiple spans and free-form and various numerical reasoning capabilities are usually required, including addition (+), subtraction (-), multiplication (x), division (/), counting, comparison, sorting, and their compositions;

In total, TAT-DQA contains 16,558 questions associated with 2,758 documents ( 3,067 document pages ) sampled from real-world financial reports.

Citation

@inproceedings{zhu2022towards,
  title={Towards complex document understanding by discrete reasoning},
  author={Zhu, Fengbin and Lei, Wenqiang and Feng, Fuli and Wang, Chao and Zhang, Haozhou and Chua, Tat-Seng},
  booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
  pages={4857--4866},
  year={2022}
}