Learning Choice Nuance for Multiple-Choice Commonsense Question Answering

Existing models for commonsense question answering (CQA) usually focus on combining pre-trained language models (PLMs) and structured knowledge graphs (KGs) for joint reasoning. However, such approaches encode a QA context (i.e., a pair of the question and a choice) separately from other choices, ineffective for explicitly capturing useful subtle differences among the choices, which results in incorrect answers in some cases. This paper proposes a novel model LNC (Learning Nuance among Choices) for addressing this problem and thus provides an improved approach to multiple-choice question answering. Specifically, LNC explicitly interacts between the text knowledge corresponding to each choice and the external KG knowledge corresponding to each choice, and removes the commonalities among similar choices, allowing the model to focus on different relevant knowledge based on the choices, thereby distinguishing semantically similar choices. Experimental results on major benchmark datasets show that LNC is competitive comparing to the baseline models.