Dataset Overview

Basic Information

• Dataset Name: Text Retrieval Conference Question Answering (TRECQA)
• Language: English (en)
• License: Unknown
• Multilinguality: Monolingual
• Size: 1K < n < 10K
• Source: Raw data
• Task Type: Text Classification
• Task ID: Multi‑class Classification
• Paper ID: trecqa
• Pretty Name: Text Retrieval Conference Question Answering

Structure

Features

• text (string): question text.
• coarse_label (categorical): coarse categories, possible values:
  • ABBR (0): abbreviation.
  • ENTY (1): entity.
  • DESC (2): description/abstract concept.
  • HUM (3): human.
  • LOC (4): location.
  • NUM (5): numeric.
• fine_label (categorical): fine categories, grouped under ABBREVIATION, ENTITY, DESCRIPTION, HUMAN, LOCATION, NUMERIC (see original for full list).

Splits

• train: 5,452 samples
• test: 500 samples

Dataset Creation

Summary

• Training set: 5,452 labeled questions
• Test set: 500 labeled questions
• Coarse categories: 6
• Fine categories: 50
• Average sentence length: 10 tokens
• Vocabulary size: 8,700

Sources

• 4,500 questions from USC (Hovy et al., 2001)
• ~500 manually created questions for rare classes
• 894 questions from TREC‑8 and TREC‑9
• 500 TREC‑10 questions as test set

Citation

@inproceedings{li-roth-2002-learning,
    title = "Learning Question Classifiers",
    author = "Li, Xin and Roth, Dan",
    booktitle = "COLING 2002",
    year = "2002",
    url = "https://www.aclweb.org/anthology/C02-1150",
}
@inproceedings{hovy-etal-2001-toward,
    title = "Toward Semantics‑Based Answer Pinpointing",
    author = "Hovy, Eduard and Gerber, Laurie and ...",
    booktitle = "First International Conference on Human Language Technology Research",
    year = "2001",
    url = "https://www.aclweb.org/anthology/H01-1069",
}