融合实体关系信息的答案选择网络中的算法分析  

Research on the Algorithms of Neural Network with Entity Relation Information for Answer Selection

作　　者：毛鹏 刘瑞芳[1] 苗航 陆树栋 李思 MAO Peng;LIU Ruifang;MIAO Hang;LU Shudong;LI Si(School of Information and Communication Engineering,Beijing University of Posts and Telecommunications,Beijing 100876)

机构地区：[1]北京邮电大学信息与通信工程学院,北京100876

出　　处：《计算机与数字工程》2021年第10期1973-1977,2032,共6页Computer & Digital Engineering

摘　　要：论文针对答案选择任务进行研究,利用深层神经网络结合外部知识库信息的方法,提出基于知识库关系信息的双向长短时记忆网络(Knowledge Based Relation-Bidirectional Long Short Term Memory,KBR-BiLSTM),引入知识库中实体信息及实体关系信息去优化基准模型中的注意力机制;并利用知识库关系信息结合上下文丰富了问答的句子编码信息,提升模型效果。在维基问答(Wiki QA)数据集和TREC QA数据集上进行对比实验,证明了KBR-BiLSTM模型的有效性。In this paper,the answer selection task is studied.This paper proposes a model named Knowledge Based Relation-Bidirectional Long Short Term Memory(KBR-BiLSTM),which introduces the relation information in the knowledge base to optimize the attention mechanism in the benchmark model.In addition,the relation of knowledge base is used.The information combined with the context enriches the sentence coding information of the question and answer.Finally,it improves the effect of KBR-BiLSTM model with multiple comparative experiments on the Wiki QA dataset and the TREC QA dataset demonstrates the validity of the proposed model.

关 键 词：深层神经网络 知识库 关系信息 注意力机制 

分 类 号：TP301.6[自动化与计算机技术—计算机系统结构]