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

CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge (ACL 2019)

Forum topic · 小凯 · 2026-07-05

Summary

CommonsenseQA, presented at ACL 2019 by Talmor, Herzig, Lourie, and Berant, is a multiple-choice question answering benchmark designed to test commonsense reasoning beyond what can be extracted from text alone. The dataset contains 12,102 questions, each with five answer options, authored by crowd workers based on ConceptNet relations, drawing on target concepts and their connected nodes to elicit prior knowledge. In a notable validation design, a BERT-large baseline pretrained on RACE achieves only about 56% accuracy, far below the roughly 89% human performance, demonstrating that the benchmark cannot be solved by reading comprehension alone and requires inference over everyday knowledge. The paper analyzes error patterns, showing that questions touching familiar people and objects are easier, and identifies ConceptNet noise as a factor. CommonsenseQA has since become a standard evaluation set for commonsense reasoning in NLP. Original paper: https://aclanthology.org/N19-1421/

CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge (ACL 2019)

Overview

This post summarizes the ACL 2019 paper CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge by Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant (Tel-Aviv University).

  • Paper page: <https://aclanthology.org/N19-1421/>
  • Venue: NAACL 2019 (ACL Anthology N19-1421)
  • Task type: multiple-choice question answering (5 options)
  • Motivation

    Modern QA systems can often answer questions by retrieving evidence from text, but many questions require prior knowledge — the kind of everyday commonsense humans acquire without explicit instruction. The authors design a benchmark where the answer cannot be inferred from a passage, forcing models to rely on learned world knowledge.

    Key points

  • Dataset: 12,102 multiple-choice questions, each with five answer options.
  • Construction: Crowd workers author questions based on ConceptNet relations, using a target concept and its neighboring nodes to create questions that probe commonsense associations.
  • Validation by difficulty: Before collecting answers, the authors test a strong baseline — BERT-large pretrained on RACE — and require it to score below a set threshold, filtering out questions solvable via textual cues alone.
  • Results

    | Model | Accuracy | |-------|----------| | Random | ~20% | | BERT-large (RACE-pretrained) | ~56% | | Human | ~89% |

    The large gap between the neural baseline and humans shows the benchmark genuinely requires commonsense reasoning, not just reading comprehension.

    Analysis

  • Questions about familiar people and objects are easier for models; abstract relations remain hard.
  • Noise in ConceptNet affects question quality and serves as an upper-bound concern for the dataset.
  • The paper proposes a knowledge-augmented two-encoder baseline, and subsequent work has built extensively on CommonsenseQA as a standard evaluation set.
  • Why it matters

    CommonsenseQA helped shift the NLP community's attention from extractive QA toward reasoning-centric evaluation, complementing datasets like SWAG, PIQA, and later BIG-bench tasks. It remains widely cited in LLM evaluation suites.

    References

  • Talmor, A., Herzig, J., Lourie, N., & Berant, J. (2019). CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge. NAACL 2019. <https://aclanthology.org/N19-1421/>

Tags

#commonsense-qa#question-answering#benchmark#nlp#conceptnet#bert#acl-2019

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/178208719