TAT-QA (Tabular And Textual QA)
Goal
The TAT-QA dataset is designed to advance hybrid question answering over financial documents. Its primary goals are:
To enable numerical reasoning over combined tabular and textual data
To support complex multi-step reasoning in financial document analysis
To benchmark hybrid QA systems for finance
Description
TAT-QA is a large-scale question answering dataset built from real-world financial reports, requiring models to jointly reason over tables and associated text to answer complex questions.
Dataset Composition
Source: Financial reports from U.S. companies (10-K, 10-Q filings)
Time Period: Recent financial filings (2018-2020)
Sample Size: 16,552 questions across 2,757 table-text pairs
Document Types: Quarterly and annual financial reports with hybrid content
Example
Context: A financial table showing quarterly revenue alongside text describing business performance
Question: “What was the total net income for the first half of 2020?”
Answer: “$78.8 million” (requires arithmetic: combining Q1 and Q2 net income)
Task Description
Hybrid Question Answering with Numerical Reasoning
Input: A question, an associated financial table, and related text passage
Output: Answer span or computed numerical value
Challenge: Requires understanding table structure, comprehending text, identifying relevant cells/spans, and performing multi-step numerical operations
Key Use Cases
Financial Analysis: Automated extraction of insights from earnings reports
Investor Decision Support: Quick answers to investor queries about company performance
Compliance and Audit: Verification of financial statement consistency
Evaluation Metrics
Exact Match (EM): Percentage of predictions exactly matching gold answers
F1-Score: Token-level F1 score for partial credit on numerical answers
References
Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, and Tat-Seng Chua. “TAT-QA: A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance.” In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL), 2021.
For dataset access, visit the official repository or HuggingFace.