基于位置和注意力联合表示的知识图谱问答

Knowledge Graph Question-Answering Based on Joint Location and Attention Representation

知识图谱是人工智能的重要组成部分,其以结构化的方式描述客观世界中的概念、实体及关系,提供了一种更优的组织、管理和理解互联网海量信息的能力。随着深度学习技术的发展,基于表示学习的知识图谱问答方法陆续出现。利用表示学习的方法实现知识图谱问答的核心目标是将问题嵌入到与三元组相同维度的表示向量空间中,通过合适的答案预测方法来匹配问题与答案。参考复数域编码的思路,构建一种基于位置和注意力联合表示的三元组表示模型Pos-Att-complex。在三元组表示部分,将词本身的特征和位置特征联合编码,并通过解码器网络进一步挖掘深层次特征,从而对三元组进行打分。在知识图谱问答部分,将问题通过RoBERTa嵌入到与三元组向量相同维度的向量空间中,并与通过关系筛选的关系集合进行向量融合。在此基础上,通过联合表示解码器为候选答案打分,以筛选出问题的答案。实验结果表明,该模型在三元组分类和多跳问答基准数据集上均能取得良好的测试结果,准确率优于GraftNet、VRN等模型。

Knowledge graph is an important part of artificial intelligence.It describes the concepts,entities,and relationships in the objective world in a structured way and provides a better ability to organize,manage,and understand the massive amount of information available on the Internet.With the development of deep learning technology,representation-learning-based knowledge graph question-answering methods have emerged.The core goal of such methods is to embed the question into the representation vector space with the same dimension as triples and match the questions and answers through appropriate answer prediction methods.Referring to the idea of complex field coding,this paper presents a triple-representation model,Pos-Att-complex,based on joint location and attention representation.In the triplet representation part,the features of the word itself and the location features are jointly encoded.The deep-seated features are further mined through the decoder network,so as to score the triplet.In the question and answer part of the knowledge graph,the question is embedded into the vector space with the same dimension as the triplet vector through RoBERTa,and the vector is fused with the relationship set filtered through the relationship.On this basis,the candidate answers are scored by the joint representation decoder to screen them out.Experimental results show that the model can achieve good test results on triple classification and multi hop question and answer benchmark datasets.Furthermore,it outperforms GraftNet,VRN,and other existing models.