结合问题-关系注意力和特征视图匹配的关系检测方法

Relation Detection Method Combining Question-Relation Attention and Feature View Matching

关键词:  关系检测|问答系统|注意力机制|信息检索|特征匹配

Key words:  relation detection|question answering system|attention mechanism|information retrieval|feature matching

摘要:

问答系统作为信息检索的一种高级形式，已成为人工智能和自然语言处理领域中一个备受关注的研究方向。本文聚焦于知识图谱问答(Knowledge Graph Question Answering，KGQA)研究中的关系检测，针对现有方法中未能较好兼顾全局语义和局部语义信息，以及复杂问题准确率不高的问题，提出了一种结合问题-关系注意力和特征视图匹配的关系检测方法。该方法从问题和知识库中提取多粒度的特征，将提取特征构造成特征对视图作为关系检测模型的输入，视图内部利用双边多视角匹配(Bilateral Multi-Perspective Matching,BiMPM)进行比较匹配，得出关系预测结果。为验证本文提出的方法，在SimpleQuestions、WebQSP数据集上，与6种主流基线方法进行对比实验，本方法的准确率分别提升3.42个和0.45个百分点。

Abstract:

As an advanced form of information retrieval,question answering system has become a research direction that has received much attention in the field of artificial intelligence and natural language processing.This article focuses on relation detection in Knowledge Graph Question Answering (KGQA) research.In view of the problems that the existing methods fail to take into account global semantic and local semantic information and the low accuracy of complex problems,a relation detection method combining question-relation attention and feature view matching is proposed.This method extracts multi-granularity features from questions and knowledge base,constructs the extracted features into feature pair view as the input of the relation detection model.Bilateral Multi-Perspective Matching (BiMPM) is used to compare and match within the view to obtain the relation prediction results.In order to verify the method proposed in this article,the accuracy of this method is improved by 3.42 and 0.45 percentage points respectively compared with the six mainstream baseline methods on the SimpleQuestions and WebQSP datasets.