G2S:基于语义块的知识图谱问答语义解析  

G2S:Semantic Segment Based Semantic Parsing for Question Answering over Knowledge Graph

作　　者：高留杰 赵文[2] 张君福 姜波 GAO Liu-jie;ZHAO Wen;ZHANG Jun-fu;JIANG Bo(School of Software and Microelectronics,Peking University,Beijing 100871,China;National Engineering Research Center for Software Engineering,Peking University,Beijing 100871,China;Beijing Beida Software Engineering Co.Ltd;Unit 31 of 96901 PLA Troops,Beijing 100085)

机构地区：[1]北京大学软件与微电子学院,北京100871 [2]北京大学软件工程国家工程研究中心,北京100871 [3]北京北大软件工程股份有限公司,北京100080 [4]96901部队31分队,北京100085

出　　处：《电子学报》2021年第6期1132-1141,共10页Acta Electronica Sinica

摘　　要：问题意图理解是知识图谱问答的主要任务之一,语义解析是当前理解问题意图的主流方法.其主要挑战是如何充分利用知识图谱上下文理解问句中的隐含实体或关系,以及时间、排序和聚合等复杂约束条件等意图.为了应对这些挑战,本文提出了一种基于语义块的知识图谱问答语义解析框架——Graph-to-Segment,框架中的语义解析模型结合了基于规则的准确度和基于深度学习的覆盖度,实现了问题到语义块序列的解析和语义查询图的构造.框架将问题意图使用基于语义块的语义查询图表示,将问题的语义解析建模为语义块序列生成任务,采用编码器-解码器神经网络模型实现问题到语义块序列的解析,然后通过语义块组装形成语义查询图.同时,结合知识图谱中的上下文信息,模型使用图神经网络学习问题的表示,改进隐含实体或关系的语义解析效果.在两个知识图谱问答数据集上的实验表明,模型性能达到了良好的效果.Question understanding is one of the important tasks of question answering over knowledge graph,where semantic parsing is the mainstream approach for understanding question utterance.The most significant challenge in this task is to understand the implicit entities,relations and the utterances of complex constraints such as time,ordinal,and aggregation in the question with the context of knowledge graph.In this paper,we propose graph-to-segment,a semantic segments based semantic parsing framework for question answering over knowledge graph.Our semantic parsing model integrates both rule-based and neural-based approaches to parse the semantic segment sequences and constructs the semantic query graphs with high accuracy and coverage.These semantic segment-based semantic query graphs,which consist of the semantic segments,are used to represent the utterance of questions.Question semantic parsing is modeled as a sequence generation task,where an encoder-decoder neural network is used to generate the semantic segments from natural language questions.Additionally,with the context information of knowledge graph,a graph neural network is used to learn the representation of questions to improve the effect of semantic parsing on implicit entities or relations.Experimental results show that our model achieves good performance on the two datasets.

关 键 词：知识图谱 问答 语义解析 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]