Beyond Alignment: Discrepancy-driven Consistency Reasoning for Commonsense Question Answering

ACL ARR 2026 January Submission2233 Authors

Keywords: commonsense QA, knowledge graph

Abstract:

Leveraging pre-trained language models (LMs) together with knowledge graphs (KGs) has become a common paradigm for commonsense question answering (CommonsenseQA), as it combines linguistic understanding with structured factual knowledge. In this setting, the core challenge lies not in the availability of external knowledge, but in selecting the candidate answer that best satisfies both linguistic semantic constraints and structured knowledge constraints. Existing language model–knowledge graph (LM–KG) approaches predominantly rely on cross-modal alignment, assuming that increased representational agreement leads to improved reasoning. However, in CommonsenseQA, semantic discrepancies between language models and knowledge graphs frequently indicate constraint violations, where linguistic plausibility conflicts with graph structure or vice versa. By smoothing out such discrepancies, alignment-centric methods obscure critical reasoning signals.
To address this limitation, we propose Semantic-Aware Reasoning Network (DCRN), which explicitly models semantic discrepancy between LMs and KGs as a reasoning signal rather than noise. DCRN leverages discrepancy-aware learning to support more reliable joint reasoning for Question Answering.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: commonsense QA, knowledge base QA

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 2233