TAT-QA

displayName: TAT-QA
license:
- MIT
paperUrl: https://arxiv.org/pdf/2105.07624v2.pdf
publishDate: "2021"
publishUrl: https://nextplusplus.github.io/TAT-QA/
publisher:
- National University of Singapore
- Tsinghua University
tags:
- Answering
taskTypes:
- Visual Question Answering

---
# 数据集介绍
## 简介
TAT-QA（用于问答的表格和文本数据集）是一个大规模的 QA 数据集，旨在促进对更复杂和现实的表格和文本数据的 QA 研究进展，特别是那些需要数值推理的数据。
TAT-QA 的独特功能包括：
给出的上下文是混合的，包括一个半结构化的表格和至少两个描述、分析或补充表格的相关段落；
这些问题是由具有丰富金融知识的人提出的，大多数是实用的；
答案形式多样，包括单跨、多跨和自由形式；
要回答这些问题，通常需要各种数值推理能力，包括加法（+）、减法（-）、乘法（x）、除法（/）、计数、比较、排序及其组合；
除了真实答案之外，如果有的话，还提供了相应的推导和尺度。
TAT-QA 总共包含 16,552 个问题，与来自真实财务报告的 2,757 个混合上下文相关联。
以下是 TAT-QA 的示例。左侧虚线框显示混合上下文。蓝色背景的行是行标题，而灰色的列是列标题。右边的实线框显示了相应的问题、答案及其比例，以及得出答案的推导。
## 引文
```
@article{zhu2021tat,
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},
journal={arXiv preprint arXiv:2105.07624},
year={2021}
}
```

## Download dataset
:modelscope-code[]{type="git"}

displayName: TAT-QA
License: MIT
Paper URL: https://arxiv.org/pdf/2105.07624v2.pdf
Publication Date: "2021"
Publication URL: https://nextplusplus.github.io/TAT-QA/
Publishers: National University of Singapore, Tsinghua University
Tags: Answering
Task Types: Visual Question Answering

---

# Dataset Introduction
## Overview
TAT-QA (Table and Text Dataset for Question Answering) is a large-scale QA dataset intended to advance research on QA over more complex and realistic tabular and textual data, particularly those requiring numerical reasoning.

Unique features of TAT-QA are as follows:
1. The provided context is a hybrid of a semi-structured table and at least two relevant paragraphs that describe, analyze, or supplement the table;
2. The questions are formulated by individuals with extensive financial expertise, with most being practical in nature;
3. Answer formats are diverse, including single-span, multi-span, and free-form answers;
4. Answering these questions typically demands various numerical reasoning capabilities, including addition (+), subtraction (-), multiplication (×), division (/), counting, comparison, sorting, and their combinations;
5. In addition to the ground-truth answers, corresponding derivations and scales (if available) are also provided.

In total, TAT-QA contains 16,552 questions linked to 2,757 mixed contexts sourced from real financial reports.

Here is an example of TAT-QA. The left dashed box displays the mixed context. Rows with blue backgrounds serve as row headers, while gray columns act as column headers. The right solid box shows the corresponding question, answer and its scale, as well as the derivation process for obtaining the answer.

## Citation

@article{zhu2021tat,
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},
journal={arXiv preprint arXiv:2105.07624},
year={2021}
}

## Download Dataset
:modelscope-code[]{type="git"}