一种面向知识图谱多跳问答的分层语义解析方法  

Hierarchical semantic parsing approach for multi-hopquestion answering on knowledge graphs

作　　者：周岩 范永胜 孙松 周粤 Zhou Yan;Fan Yongsheng;Sun Song;Zhou Yue(College of Computer&Information Science,Chongqing Normal University,Chongqing 401331,China)

机构地区：[1]重庆师范大学计算机与信息科学学院,重庆401331

出　　处：《计算机应用研究》2025年第3期714-719,共6页Application Research of Computers

基　　金：重庆市教育委员会科学技术研究项目(KJQN202300547);浙江大学工业控制技术国家重点实验室开放研究项目(ICT2023B40)。

摘　　要：在知识图谱多跳问答任务中,问题的复杂语义往往难以被完全理解,这常导致回答的准确性低于预期。为此,提出了一种名为HL-GPT(hierarchical parsing and logical reasoning GPT)的新型框架。该框架利用大语言模型的理解和生成能力,通过分层解析复杂语义及构建逻辑推理路径,以提升问答任务的准确性。研究方法包括三个阶段:首先,通过大语言模型从问题的不同层次中解析出关键实体和多层关系,并将这些信息转换为逻辑形式;其次,将这些逻辑形式与知识图谱中的数据进行映射,并逐步检索相关实体以构建逻辑推理路径;最后,利用大语言模型固有的推理能力,整合问题和逻辑路径以生成准确的答案。本框架在MetaQA、COKG-DATA、AeroQA和NLPCC-MH四个数据集上进行实验,实验结果表明,HL-GPT相比基线模型有明显的性能提升。In multi-hop question answering tasks over knowledge graphs,the complex semantics of questions often remain inadequately understood,leading to suboptimal accuracy in answers.To address this challenge,this paper introduced a novel framework named HL-GPT.This framework exploited the comprehension and generation capabilities of large language models to enhance answer accuracy through hierarchical semantic parsing and logical reasoning path construction.The method encompassed three stages:initially,a large language model parsed key entities and multi-layer relationships from various levels of the question and converted this information into logical forms.Subsequently,it mapped these logical forms to data within the knowledge graph and incrementally retrieves relevant entities to construct a logical reasoning path.Finally,it utilized the inherent reasoning capabilities of the large language model to integrate the question and logical path,generating accurate answers.Experimental results on the MetaQA,COKG-DATA,AeroQA and NLPCC-MH datasets demonstrate significant perfor-mance improvements of HL-GPT over baseline models.

关 键 词：大语言模型 知识图谱 多跳问答 语义解析 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]