Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -6,6 +6,42 @@ language:
|
|
6 |
- en
|
7 |
tags:
|
8 |
- finance
|
|
|
|
|
|
|
9 |
size_categories:
|
10 |
- 10K<n<100K
|
11 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
6 |
- en
|
7 |
tags:
|
8 |
- finance
|
9 |
+
- table-text
|
10 |
+
- visual-document-QA
|
11 |
+
- numerical-reasoning
|
12 |
size_categories:
|
13 |
- 10K<n<100K
|
14 |
+
---
|
15 |
+
|
16 |
+
# TAT-DQA
|
17 |
+
|
18 |
+
- [**Project Page**](https://nextplusplus.github.io/TAT-DQA/)
|
19 |
+
- [**Paper - MM 22**](https://dl.acm.org/doi/abs/10.1145/3503161.3548422)
|
20 |
+
- [**Paper - Arxiv**](https://arxiv.org/abs/2207.11871)
|
21 |
+
- [**Github**](https://github.com/NExTplusplus/TAT-DQA)
|
22 |
+
- [**Leaderboard**](https://nextplusplus.github.io/TAT-DQA/#leaderboard)
|
23 |
+
|
24 |
+
|
25 |
+
**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.
|
26 |
+
|
27 |
+
The unique features of TAT-DQA include:
|
28 |
+
|
29 |
+
- The documents in TAT-DQA dataset are sampled from real-world high-quality financial reports and each document contains both tabular and textual data;
|
30 |
+
- The average number of words of each document in TAT-DQA is around 550, which is significantly larger than all existing Document VQA datasets.
|
31 |
+
- Around 85% of the documents in the dataset have only one page while 15% has multiple pages.
|
32 |
+
- 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;
|
33 |
+
|
34 |
+
|
35 |
+
In total, TAT-DQA contains 16,558 questions associated with 2,758 documents ( 3,067 document pages ) sampled from real-world financial reports.
|
36 |
+
|
37 |
+
|
38 |
+
## Citation
|
39 |
+
```python
|
40 |
+
@inproceedings{zhu2022towards,
|
41 |
+
title={Towards complex document understanding by discrete reasoning},
|
42 |
+
author={Zhu, Fengbin and Lei, Wenqiang and Feng, Fuli and Wang, Chao and Zhang, Haozhou and Chua, Tat-Seng},
|
43 |
+
booktitle={Proceedings of the 30th ACM International Conference on Multimedia},
|
44 |
+
pages={4857--4866},
|
45 |
+
year={2022}
|
46 |
+
}
|
47 |
+
```
|