UDA Paper Tab QA (orgrctera/uda_paper_tab_qa_qa)

Overview

This dataset is the Paper Tab QA slice of the UDA (Unstructured Document Analysis) benchmark: 393 question–answer instances over academic research papers (PDF), packaged for question answering evaluation in RAG and document-analysis pipelines.

UDA (Hui et al., NeurIPS 2024 Datasets & Benchmarks) is a suite for Retrieval-Augmented Generation over messy, real-world documents. Sources are kept in original formats (e.g. PDF, HTML) so that parsing, chunking, and retrieval are evaluated alongside generation. The benchmark spans finance, academia, and Wikipedia-style knowledge bases, with 2,965 documents and 29,590 expert-annotated Q&A pairs in total.

Within the academia track, PaperTab targets questions whose answers are grounded in tabular and figure-related content inside scholarly PDFs (as opposed to narrative-only PaperText). The UDA paper reports 307 PaperTab documents and 393 Q&A pairs for this subset; this Hub release exposes 393 rows in the default split, aligned with that configuration.

Task

• Task type: Question answering (QA) for Paper Tab QA — answer natural-language questions about information that appears in tables, statistics, and related structured fragments within academic papers.
• Input: A question string (input) about the paper (e.g. reported metrics, experimental settings, language pairs, or results tied to tabular evidence).
• Supervision / reference: expected_output is a JSON string with gold answers (possibly multiple references with abstractive, extractive, or other answer types) and evidence spans (e.g. raw_evidence, highlighted_evidence, including lines such as FLOAT SELECTED: Table … that point to table regions). metadata records UDA identifiers (sub_benchmark: paper_tab_qa).

Systems are typically evaluated on answer correctness (exact or semantic match to reference answers) and grounding (whether retrieved or generated content is supported by the same tables and passages as the gold evidence), consistent with UDA’s emphasis on end-to-end document analysis.

Background

Paper Tab QA in UDA

Academic PDFs mix body text, equations, and tables/figures. Paper Tab QA instances are designed so that answering well requires attending to tabular and semi-structured material—not only adjacent paragraphs. That makes the subset a useful stress test for layout-aware parsing, table retrieval, and multi-evidence reasoning in scientific domains.

UDA benchmark (suite)

UDA revisits LLM- and RAG-based solutions for document analysis across domains and query types. The authors highlight that data parsing and retrieval design strongly affect answer quality when documents are long and unstructured. The suite and code are public; see references below.

Data fields

Column	Type	Description
input	string	Natural-language question about the paper.
expected_output	string	JSON string with answers (list of { "answer", "type" }) and evidence (lists of supporting strings, often including table captions and FLOAT SELECTED: … markers).
metadata	struct	benchmark_name (uda_paper_tab_qa), benchmark_type (uda), split, sub_benchmark (paper_tab_qa), and value (JSON string with label_key, label_file, q_uid).

Splits: Single split default with 393 examples.

Examples

Illustrative rows from this release (expected_output is shown formatted for readability; strings may be truncated in documentation).

Example 1 — language pairs (abstractive / extractive answers)

• 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": ["..."]
    }
  ]
}

Example 2 — reported accuracy (abstractive answers, table evidence)

• input: What accuracy does the proposed system achieve?
• expected_output (excerpt):

{
  "answers": [
    {
      "answer": "F1 scores of 85.99 on the DL-PS data,  75.15 on the EC-MT data and 71.53 on the EC-UQ data ",
      "type": "abstractive"
    },
    {
      "answer": "F1 of 85.99 on the DL-PS dataset (dialog domain); 75.15 on EC-MT and 71.53 on EC-UQ (e-commerce domain)",
      "type": "abstractive"
    }
  ],
  "evidence": [
    {
      "raw_evidence": [
        "FLOAT SELECTED: Table 2: Main results on the DL-PS data.",
        "FLOAT SELECTED: Table 3: Main results on the EC-MT and EC-UQ datasets."
      ],
      "highlighted_evidence": [
        "FLOAT SELECTED: Table 2: Main results on the DL-PS data.",
        "FLOAT SELECTED: Table 3: Main results on the EC-MT and EC-UQ datasets."
      ]
    }
  ]
}

metadata.value (example):

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

References

UDA benchmark (primary reference)

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: The use of Retrieval-Augmented Generation (RAG) has improved Large Language Models (LLMs) in collaborating with external data, yet significant challenges exist in real-world scenarios. In areas such as academic literature and finance question answering, data are often found in raw text and tables in HTML or PDF formats, which can be lengthy and highly unstructured. In this paper, we introduce a benchmark suite, namely Unstructured Document Analysis (UDA), that involves 2,965 real-world documents and 29,590 expert-annotated Q&A pairs. We revisit popular LLM- and RAG-based solutions for document analysis and evaluate the design choices and answer qualities across multiple document domains and diverse query types. Our evaluation yields interesting findings and highlights the importance of data parsing and retrieval.

• arXiv: https://arxiv.org/abs/2406.15187
• DOI: https://doi.org/10.48550/arXiv.2406.15187
• NeurIPS proceedings: Datasets and Benchmarks Track abstract
• Code & resources: https://github.com/qinchuanhui/UDA-Benchmark

Related Hub resources

• Aggregated UDA QA (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 underlying paper data licenses and terms. Verify conditions for your use case before redistribution or commercial use.