利用知识图谱的多跳可解释问答  

Multi-hop Interpretable Question Answering Using Knowledge Graph

作　　者：叶蕾[1] 张宇迪 杨旭华[1] YE Lei;ZHANG Yudi;YANG Xuhua(College of Computer Science and Technology,Zhejiang University of Technology,Hangzhou 310023,China)

机构地区：[1]浙江工业大学计算机科学与技术学院,杭州310023

出　　处：《小型微型计算机系统》2024年第8期1869-1877,共9页Journal of Chinese Computer Systems

基　　金：国家自然科学基金项目(62176236)资助.

摘　　要：基于知识图谱的多跳问答需要分析和理解自然语言问题并在知识图谱的实体和关系上经过多次推理获取答案,是自然语言处理的重要研究领域.现有的模型一般通过知识图谱与问题嵌入,利用神经网络推断答案;或使用一阶逻辑规则结合概率方法预测答案;前者缺乏可解释性,后者在复杂问题中性能欠佳.为解决上述问题,本文提出一种基于知识图谱的多跳可解释问答方法(MIQA),它通过在实体间的多次跳跃推理来获取答案.MIQA首先使用BERT预训练模型获取自然语言问题表征向量以及问题分词后的词向量矩阵,在每一跳中,结合问题向量提取问题当前时刻的特征向量,根据特征向量的分类结果计算下一跳的关系分数和实体分数,多次跳跃后,综合分数最高的实体被作为答案,而获取该答案所对应的路径为推理路径.该方法推理准确率高,同时具有明显的可解释性.在MetaQA、WebQuestionsSP、ComplexWebQuestions这3个数据集上,通过和其他8个知名算法相比较,仿真结果表明MIQA性能优异,达到了当前的SOTA.Multi-hop question answering based on knowledge graphs needs to analyze and understand natural language questions and obtain answers through multiple reasoning on the entities and relationships of knowledge graphs,which is an important research field of natural language processing.Existing models generally use knowledge graphs and question embeddings to infer answers using neural networks;or use first-order logic rules combined with probabilistic methods to predict answers;the former lacks interpretability,and the latter performs poorly in complex problems.In order to solve the above problems,this paper proposes a multi-hop interpretable question answering method(MIQA)based on knowledge graph,which obtains answers by reasoning through multiple jumps between entities.MIQA first uses the BERT pre-training model to obtain the natural language question representation vector and the word vector matrix after the question word segmentation.In each hop,it combines the question vector to extract the feature vector at the current moment of the question,and calculates the next hop based on the classification result of the feature vector.The relationship score and entity score,after multiple jumps,the entity with the highest comprehensive score is taken as the answer,and the path corresponding to the answer is the inference path.The method has high inference accuracy and obvious interpretability.On the three data sets of MetaQA,WebQuestionsSP,and ComplexWebQuestions,compared with other 6 well-known algorithms,the simulation results show that MIQA has excellent performance,reaching the current SOTA.

关 键 词：知识图谱 多跳问答 可解释性 特征抽取 注意力机制 

分 类 号：TP182[自动化与计算机技术—控制理论与控制工程]