关键词: 常识性问答, 多项选择问答, 知识融合, 知识图谱, 语言模型

Abstract: Knowledge graphs have demonstrated significant effectiveness in commonsense question answering.Existing methods typically utilize entities from the question to retrieve local subgraphs from the knowledge graph(KG),which are then encoded using graph neural networks(GNN).Subsequently,the GNN-encoded results are combined with language models(LMs) to infer answers and answer the questions.However,commonsense question answering systems using GNNs and LMs face two challenges:1) how to efficiently extract subgraphs from the knowledge graph,effectively represent and utilize their knowledge and structural information; 2) how to achieve deep integration and joint reasoning of the question context and subgraph knowledge.This paper proposes a graph-text integrating model for commonsense question answering(Graph-Text Integrating for Commonsense Question Answering,GTICQA).The model initially refines key entities by filtering through an external dictionary,achieving pruning of the knowledge subgraph,and then separately encodes the question context using an LM and the refined knowledge subgraph using a GNN encoder.Additionally,during the subgraph encoding process,a novel k-sparse attention mechanism is introduced to enhance the extraction of global features from the subgraph and suppress noise.Finally,a knowledge fusion method that includes fine-grained bimodal interaction fusion layers and mean interaction fusion layers is used to deeply integrate and dynamically update the two knowledge representations.The GTICQA model is evaluated on three datasets:CommonsenseQA,OpenBookQA,and MedQA-USMLE,achieving accuracy rates of 79.12%,72.20%,and 39.40%,respectively,surpassing the current best methods,demonstrating the model's advantage in handling commonsense question answering.

Key words: Commonsense QA, Multiple choice QA, Knowledge integration, Knowledge graph, Language model