基于层次结构图的多跳知识图谱问答模型
Multi-hop Knowledge Base Question Answering Model Based on Hierarchical Structure Graph
刘昀抒 1申彦明 1齐恒 1尹宝才2
作者信息
- 1. 大连理工大学计算机科学与技术学院,辽宁 大连 116024
- 2. 北京工业大学人工智能与自动化学院,北京 100124
- 折叠
摘要
知识图谱问答(KBQA)旨在理解用户的自然语言问句,在结构化的知识图谱中通过检索、推理等手段来获取答案实体.近年来,多跳KBQA备受关注,然而,复杂问句中通常存在多个关系意图,已有KBQA方法大多忽视了推理关系链的关系顺序问题.为此,提出一种基于层次结构图的多跳知识图谱问答模型(HSG-KBQA),建模自然语言问句的关系层次顺序,指导模型在每个推理步选择合理的关系意图.设计一种层次结构图,显式地体现问句中关系的层次距离,利用LSTM-BiGCN编码层将词语间的依存信息编码到问句中;提出虚拟节点的概念,利用图池化技术过滤不重要的节点,学习推理过程中知识图谱的状态;设计基于注意力机制和层次权重的解码器来优化指令生成,使推理指令更匹配问句中的关系链顺序.实验结果表明,HSG-KBQA在WebQuestionsSP数据集上取得了71.3%的Hits@1分数,在PathQuestions数据集上取得了97.3%(PQ-2H)和89.7%(PQ-3H)的Hits@1分数,均优于对照基准模型,表明HSG-KBQA模型在KBQA任务中具有更好的性能.
Abstract
Knowledge Base Question Answering(KBQA)aims to interpret natural language questions and obtain answer entities via retrieval and inference in a Knowledge Graph(KG).Recently,considerable attention has been given to multi-hop KBQA,which faces a significant,yet often overlooked challenge:complex questions typically contain multiple relational intentions with an inferential order that is frequently ignored.To address this challenge,this paper proposes a multi-hop KBQA model based on a Hierarchical Structure Graph(HSG-KBQA).This model is designed to recognize and guide the relational hierarchical order of natural language questions,enabling the selection of appropriate relational intents at each inference step.The proposed hierarchical structure graph explicitly represents the relational hierarchical order in questions and encodes the dependency information into the question using an LSTM-BiGCN encoding layer.Furthermore,the concept of a virtual node is introduced to learn the state of the KG during inference utilizing graph pooling techniques.Additionally,an attention mechanism-based decoder with hierarchical weights is presented,aiming to optimize instruction generation.This ensures that each hop instruction aligns more accurately with the relational chain order in the question.Experimental results show that the HSG-KBQA model demonstrates superior performance,achieving Hits@1 scores of 71.3%on the WebQuestionsSP dataset and 97.3%(PQ-2H)and 89.7%(PQ-3H)on the PathQuestions dataset.These results are superior to those of the benchmark model,indicating the enhanced capability of the proposed model in handling the KBQA task.
关键词
知识图谱问答/问答系统/多跳问答/图神经网络/动态推理Key words
Knowledge Base Question Answering(KBQA)/question answering system/multi-hop question answering/graph neural network/dynamic reasoning引用本文复制引用
基金项目
大连市科技创新基金(2022JJ12SN052)
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据