Relation Extraction as Text Matching: A Scheme for Multi-hop Knowledge Base Question Answering

Abstract

AbstractChinese Knowledge Base Question Answering (CKBQA) aims to predict answers for Chinese natural language questions by reasoning over facts in the knowledge base. In recent years, more attention has been paid to complex problems in KBQA, including multi-hop questions, multi-entity constrained questions, and questions with filtering or ordering. To this end, in this paper, we propose a KBQA system to generate logic forms and retrieve answers for different types of complex questions. For the single-hop and multi-hop questions, which account for the largest proportion among all questions, we propose a novel path construction method that converts the path construction task into a hop-by-hop relation extraction task via text matching. Our method combines the advantages of both the relation extraction and path ranking methods, which can focus on the order of entity-relations as well as alleviate the problem of exponential growth of candidate paths with the number of hops. Our proposed method achieves the averaged F1-score of 75.70% on the final leaderboard of CCKS-2022 CKBQA task.KeywordsKnowledge base question answeringMulti-hop relation extractionSemantic parsing