English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

QASPER: A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers

Forum topic · 小凯 · 2026-07-05

Summary

QASPER is a question answering dataset introduced by researchers from the Allen Institute for AI (Pradeep Dasigi, Kyle Lo, Iz Beltagy, Matt Gardner) and others (Arman Cohan, Noah A. Smith) in a May 2021 arXiv paper (arXiv:2105.03011). The dataset contains 5,049 natural language information-seeking questions posed by NLP practitioners, each anchored in one of 1,585 real research papers. Answers can be extractive (spans of the paper), abstractive (free-text derived from the paper), yes/no, or indicate that no answer exists in the paper. QASPER is designed to test systems' ability to answer questions about full scientific documents rather than short passages, making it a benchmark for scientific literature understanding and scholarly QA. The paper establishes baselines using extractive and abstractive QA models adapted to long documents, showing that performance lags well behind human performance and leaves substantial room for improvement. It remains a widely used evaluation resource for research on scientific document comprehension, retrieval-augmented generation, and long-context language models.

QASPER: A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers

Overview

QASPER (Question Answering on Scientific Papers) was presented in the paper *"A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers"* (May 2021, arXiv:2105.03011), authored by Pradeep Dasigi, Kyle Lo, Iz Beltagy, Arman Cohan, Noah A. Smith, and Matt Gardner (Allen Institute for AI; University of Washington; Yale University).

Key Points

  • Motivation: Most QA benchmarks operate over short passages, yet real information needs—especially in science—require reasoning over entire research papers. QASPER targets questions written by NLP practitioners who actually read the papers.
  • Dataset composition: 5,049 questions, each tied to one of 1,585 papers drawn from the NLP domain (Semantic Scholar corpus).
  • Answer types:
  • *Extractive*: a span copied from the paper
  • *Abstractive*: a free-form answer synthesized from the paper's content
  • *Yes/no*: boolean questions
  • *Unanswerable*: questions for which the paper contains no answer (roughly a quarter of the data)
  • Evidence: Answers are grounded in specific evidence-containing sections of the paper, supporting evidence identification as an auxiliary task.
  • Methods and Baselines

    The authors adapt existing extractive and abstractive QA architectures to operate on full papers, using hierarchical/long-input encoders to handle document length. The paper formalizes tasks for answer extraction, answer generation, yes/no decision, and unanswerability detection.

    Results and Takeaways

  • Baseline models perform substantially worse than humans on both answer extraction and generation, indicating that scientific-document QA is far from solved.
  • Performance degrades when models must reason across the full paper rather than oracle evidence.
  • QASPER has since become a standard benchmark for scientific literature understanding, retrieval-augmented generation (RAG), and long-context language models.
  • Relevance

    QASPER sits at the intersection of information retrieval, scientific NLP, and question answering. It complements benchmarks like Natural Questions and MS MARCO by focusing on full-length scholarly documents, making it a natural evaluation target for systems that retrieve, read, and reason over research papers.

    Reference

  • Dasigi, P., Lo, K., Beltagy, I., Cohan, A., Smith, N. A., & Gardner, M. (2021). *A Dataset of Information-Seeking Questions and Answers Anchored in Research Papers*. arXiv:2105.03011. https://arxiv.org/abs/2105.03011

Tags

#qasper#question-answering#scientific-nlp#datasets#information-retrieval#long-document-understanding#rag

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178208676