Context-Paraphrase Enhanced Commonsense Question Answering

Anonymous

Abstract: Commonsense question answering (CQA) generally means that the machine use the mastered commonsense to answer questions without relevant background material, which is a challenging task in natural language processing. Many prior methods mainly retrieve question related evidences from the structured knowledge base as the background material of the question, while the extracted evidence is generally described through the entities and the relationship between the entities, making it difficult for the machine to understand the meaning of the evidence completely. In this paper, we integrate the paraphrase in WordNet and Wiktionary into the evidence extraction process and machine reading comprehension (MRC) model, and propose a context-paraphrase enhanced commonsense question answering method. Specifically, the context-paraphrase obtained by WordNet and Wiktionary is first incorporated into the construction process of the heterogeneous graph, and the question related triple is extracted based on the heterogeneous graph, the triple is converted to triple-text based on a relational template. Then, the triple-text is used as the context of the question to establish an association graph containing the relationship between the context entities and the paraphrases. We further integrate the association graph into the MRC model to better guide the model to answer. Experimental results on CommonsenseQA and OpenBookQA show that context-paraphrase is effective in improving the answer accuracy of the MRC model.