UDA TAT-QA — Question Answering (orgrctera/uda_tat_qa_qa)

Overview

This release is the Question Answering (QA) packaging of the TAT-QA slice from the UDA (Unstructured Document Analysis) benchmark: 14,703 labeled question–answer instances grounded in real financial reports where evidence mixes tables and narrative text.

UDA (Hui et al., NeurIPS 2024) evaluates retrieval-augmented generation and LLM-based document analysis on original PDF/HTML documents. Its finance track includes TatHybrid—TAT-QA–aligned annotations spanning 170 documents and 14,703 Q&A pairs in the configuration described in the UDA paper.

TAT-QA (Zhu et al., ACL 2021)—Tabular And Textual dataset for Question Answering—introduces large-scale QA over hybrid contexts: each instance pairs a semi-structured table with multiple paragraphs from the same report. Questions demand numerical reasoning (arithmetic, counting, comparison, sorting, and compositions), span extraction, and careful handling of scales and answer types. The original work reports a large gap between strong neural baselines and human performance, highlighting the difficulty of joint table–text reasoning in finance.

In this Hub dataset, each row is a QA supervision instance: the model’s goal is to answer the question using the report (typically after retrieval and/or layout parsing in a full pipeline). Gold answers, supporting facts, and optional derivation strings are provided in structured form for training and evaluation.

Task: Question Answering for TAT-QA

• Task type: Question Answering on TAT-QA–style financial QA: given a natural-language question about figures, policies, or relationships in a disclosure, systems should produce the correct answer with evidence grounded in table and/or text.
• Input: input — the question string.
• Supervision: expected_output — JSON with gold answers (value, answer_type, scale), facts (strings pointing to relevant evidence), and optional derivation (e.g. arithmetic trace). metadata ties each row to UDA/TAT-QA identifiers (sub_benchmark: tat_qa).

Evaluation follows TAT-QA conventions (e.g. span matching, numeric checks, reasoning alignment) and can be combined with retrieval or parsing metrics when embedded in a RAG or document AI stack.

Background

TAT-QA (source task)

TAT-QA is built from real-world financial reports. Hybrid table + text contexts require models to align numbers across modalities, perform multi-step operations, and output answers as spans, numbers, or computed values. The benchmark popularized TAGOP-style pipelines (tagging cells/spans then symbolic reasoning) and remains a standard reference for financial hybrid QA.

UDA benchmark (suite)

UDA provides 2,965 real-world documents and 29,590 expert-annotated Q&A pairs across domains, stressing parsing, chunking, and retrieval alongside generation. The TatHybrid subset corresponds to this dataset’s 14,703 examples on the default split.

Data fields

Column	Type	Description
input	string	Natural-language question about the report.
expected_output	string	JSON: answers (with answer, answer_type, scale), facts, derivation.
metadata	struct	benchmark_name (uda_tat_qa), benchmark_type (uda), split, sub_benchmark (tat_qa), value (JSON string with ids such as label_key, label_file, q_uid, doc_page_uid).

Splits: default — 14,703 examples.

Examples

Example 1 — span answer (benefits)

• input: What benefits are provided by the company to qualifying domestic retirees and their eligible dependents?
• expected_output:

{
  "answers": {
    "answer": [
      "certain postretirement health care and life insurance benefits"
    ],
    "answer_type": "span",
    "scale": ""
  },
  "facts": [
    "certain postretirement health care and life insurance benefits"
  ],
  "derivation": ""
}

Example 2 — arithmetic answer (pension interest cost)

• input: What is the change in Interest cost on benefit obligation for pension benefits from December 31, 2018 and 2019?
• expected_output:

{
  "answers": {
    "answer": 129,
    "answer_type": "arithmetic",
    "scale": ""
  },
  "facts": ["1,673", "1,802"],
  "derivation": "1,802-1,673"
}

metadata.value (illustrative):

{
  "label_key": "overseas-shipholding-group-inc_2019",
  "label_file": "tat_qa",
  "q_uid": "bbdcf6da614f34fdb63995661c81613f",
  "doc_page_uid": "2ef48dc98e756493f097d01acf8101a2"
}

References

TAT-QA — primary data lineage

Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, Tat-Seng Chua. TAT-QA: A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance. Proceedings of ACL 2021, pages 3086–3101.

Abstract (from the publication): The paper introduces TAT-QA, with 16,552 questions over 2,757 hybrid contexts from financial reports—each context combines a table with related paragraphs. Questions require diverse numerical reasoning; answers may be spans or computed values. The authors propose TAGOP and report 58.0 F1 versus 90.8 F1 for human experts, motivating continued research on hybrid financial QA.

• ACL Anthology: https://aclanthology.org/2021.acl-long.254/
• arXiv: https://arxiv.org/abs/2105.07624
• Project page: https://nextplusplus.github.io/TAT-QA/
• Reference HF dataset: next-tat/TAT-QA
• Code: NExTplusplus/TAT-QA

UDA — benchmark suite (TatHybrid / finance)

Yulong Hui, Yao Lu, Huanchen Zhang. UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-world Document Analysis. NeurIPS 2024 (Datasets and Benchmarks Track).

Abstract (summary): UDA aggregates thousands of documents and tens of thousands of expert Q&A pairs across finance, academia, and Wikipedia-style settings, with documents kept in original formats to evaluate end-to-end RAG and analysis—including retrieval and layout challenges.

• arXiv: https://arxiv.org/abs/2406.15187
• NeurIPS proceedings: https://proceedings.neurips.cc/paper_files/paper/2024/hash/7c06759d1a8567f087b02e8589454917-Abstract-Datasets_and_Benchmarks_Track.html
• Repository: https://github.com/qinchuanhui/UDA-Benchmark
• Related Hub aggregation: qinchuanhui/UDA-QA

Citation

If you use this dataset, cite TAT-QA and UDA, and reference this dataset record as appropriate:

@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)},
  year = {2021},
  pages = {3086--3101}
}

@article{hui2024uda,
  title   = {UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-world Document Analysis},
  author  = {Hui, Yulong and Lu, Yao and Zhang, Huanchen},
  journal = {arXiv preprint arXiv:2406.15187},
  year    = {2024}
}

License

Use this dataset in accordance with the TAT-QA and UDA licenses and terms for the underlying benchmarks. Confirm suitability for your use case (including commercial or redistribution constraints) via the original project pages and licenses.