UDA Paper Tab QA (orgrctera/uda_paper_tab_qa)

Overview

This dataset is the Paper Tab slice of the UDA (Unstructured Document Analysis) benchmark: 393 question–answer instances over academic research papers, packaged for retrieval-oriented evaluation in RAG pipelines. Questions are designed so that answers are grounded in tabular and semi-structured content (figures, result tables, statistics blocks) alongside surrounding paper text—mirroring how analysts read papers in practice.

UDA is a benchmark suite for Retrieval-Augmented Generation (RAG) over messy, real-world documents (notably PDF and HTML), where evidence mixes long-form prose, tables, and layout-specific structure. The full suite comprises 2,965 real-world documents and 29,590 expert-annotated Q&A pairs across finance, academia, and knowledge-base settings. Within the academia track, Paper Tab complements Paper Text by focusing on table-centric QA rather than purely narrative spans.

In the UDA paper’s reported configuration, the Paper Tab subset covers on the order of 307 source PDFs with 393 labeled Q&A pairs (multi-page papers, long contexts; the paper reports roughly 6.1k words and 11 pages per document on average—statistics refer to the original benchmark release).

This Hub release stores one row per retrieval task instance: systems must retrieve the right regions (tables, captions, related paragraphs) and ground answers in that evidence—consistent with UDA’s emphasis on parsing, chunking, and retrieval as first-class problems, not only generation.

Task

• Task type: Retrieval (within a RAG / document-analysis pipeline) for Paper Tab QA—question answering over scientific papers where tables and structured displays carry key evidence.
• Input: A natural-language question (input) about results, comparisons, settings, or quantities presented in tabular or semi-tabular form in the paper.
• Supervision / reference: expected_output is a JSON string with answers (possibly multiple, with types such as extractive vs. abstractive consolidations) and evidence objects linking to raw and highlighted supporting strings (including table references where applicable). metadata records UDA identifiers (sub_benchmark: paper_tab_qa).

Evaluation typically combines retrieval quality (whether the correct table rows, captions, or adjacent text are retrieved) with answer correctness (string overlap, numeric match, or protocol-specific scoring), following UDA and the original benchmark tooling.

Background

Why “Paper Tab” in UDA?

Scientific PDFs bundle narrative argument with dense numeric and structural evidence in tables and figures. Models that only read running text often miss the answer; systems must align questions to the right table, read headers and units, and sometimes integrate caption text and body paragraphs. Paper Tab QA encodes that difficulty inside a broader RAG benchmark: documents stay in realistic, unstructured form, stressing document IR and layout-aware parsing upstream of any LLM.

UDA benchmark (suite containing this slice)

UDA revisits LLM- and RAG-based document analysis across domains using thousands of real-world documents and tens of thousands of expert-annotated Q&A pairs, with sources kept in original formats to stress parsing and retrieval as well as generation. Subsets span finance (e.g. hybrid table–text reports) and academia (paper text vs. paper tables), plus knowledge-base-style tasks in the full suite.

Data fields

Column	Type	Description
input	string	Question text posed over the paper.
expected_output	string	JSON with answers (list of objects with answer and type, e.g. extractive / abstractive) and evidence (list of objects with raw_evidence and highlighted_evidence string lists, often pointing at table lines and surrounding text).
metadata	struct	benchmark_name (uda_paper_tab_qa), benchmark_type (uda), split, sub_benchmark (paper_tab_qa), and value (JSON string with identifiers such as label_key—often an arXiv-style id—label_file, q_uid).

Splits: Single split default with 393 examples.

Examples

The following rows illustrate the schema (expected_output is formatted for readability; strings may be truncated).

Example 1 — language pairs (table + multi-paragraph evidence)

• input: what language pairs are explored?
• expected_output (excerpt):

{
  "answers": [
    {
      "answer": "De-En, En-Fr, Fr-En, En-Es, Ro-En, En-De, Ar-En, En-Ru",
      "type": "abstractive"
    },
    {
      "answer": "French-English-Spanish (Fr-En-Es), German-English-French (De-En-Fr) and Romanian-English-German (Ro-En-De), Arabic (Ar), Spanish (Es), and Russian (Ru), and mutual translation between themselves constitutes six zero-shot translation",
      "type": "extractive"
    }
  ],
  "evidence": [
    {
      "raw_evidence": [
        "For MultiUN corpus, we use four languages: English (En) is set as the pivot language...",
        "The statistics of Europarl and MultiUN corpora are summarized in Table TABREF18...",
        "FLOAT SELECTED: Table 1: Data Statistics."
      ],
      "highlighted_evidence": ["..."]
    }
  ]
}

• metadata.value (parsed):

{
  "label_key": "1912.01214",
  "label_file": "paper_tab_qa",
  "q_uid": "5eda469a8a77f028d0c5f1acd296111085614537"
}

Example 2 — quantitative improvement (table-centric)

• input: By how much is performance improved with multimodality?
• expected_output (excerpt):

{
  "answers": [
    {
      "answer": "by 2.3-6.8 points in f1 score for intent recognition and 0.8-3.5 for slot filling",
      "type": "abstractive"
    },
    {
      "answer": "F1 score increased from 0.89 to 0.92",
      "type": "abstractive"
    }
  ],
  "evidence": [
    {
      "raw_evidence": [
        "FLOAT SELECTED: Table 1: Speech Embeddings Experiments: Precision/Recall/F1-scores (%) of NLU Models"
      ],
      "highlighted_evidence": [
        "FLOAT SELECTED: Table 1: Speech Embeddings Experiments: Precision/Recall/F1-scores (%) of NLU Models"
      ]
    }
  ]
}

References

UDA benchmark (primary reference for Paper Tab QA)

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 (short): Introduces UDA with thousands of real-world documents and tens of thousands of expert-annotated Q&A pairs spanning finance, academia, and knowledge bases; evaluates LLM- and RAG-based pipelines and highlights the role of data parsing and retrieval in end-to-end document analysis.
• 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
• Code & resources: https://github.com/qinchuanhui/UDA-Benchmark

Related Hub resources

• Aggregated UDA QA release (reference): qinchuanhui/UDA-QA

Citation

If you use this dataset, please cite UDA (and this dataset record as appropriate):

@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 compliance with the UDA benchmark and upstream data licenses. Verify conditions for your use case (including redistribution or commercial use) against the official UDA repository and documentation.