融合语言知识与深度学习的文本蕴含识别  

Text Entailment Recognition Based on Integration of Language Knowledge and Deep Learning

作　　者：郑德权[1] 于凤[1] 王贺伟 ZHENG Dequan;YU Feng;WANG Hewei(School of Computer and Information Engineering,Harbin University of Commerce,Harbin 150028,China;School of Computer Science and Technology,Harbin Institute of Technology,Harbin 150001,China)

机构地区：[1]哈尔滨商业大学计算机与信息工程学院,哈尔滨150028 [2]哈尔滨工业大学计算机科学与技术学院,哈尔滨150001

出　　处：《计算机工程与应用》2020年第24期157-163,共7页Computer Engineering and Applications

基　　金：黑龙江省自然科学基金(No.F2017020);国家重点研发计划(No.2017YFB1002102)。

摘　　要：文本蕴含技术在自然语言处理中得到了广泛应用,但存在词对推理能力差的问题(例如,句对中出现反义词对无法判断反义关系等)。重点研究了词对知识向量的获取问题,包括融合多特征及有监督的词对关系向量获取、采用TransR的词对关系表示获取、反义词向量表示获取等三种方法,并将知识向量引入到文本蕴含识别模型中的词对齐和注意力机制部分。有关实验表明,上述方法相比经典模型有了较大的提升。Text entailment technology has been widely used in natural language processing,but there are some problems such as the poor reasoning ability of word pairs(for example,there is the antonym pairs in sentence pairs,but can’t judge the antonym relationship,etc.).This paper focuses on the acquisition of knowledge vector from words,including acquire the word pairs relation vector by integration of multi feature and supervised method,acquire word pairs relation expression using TransR tools,and acquire antonym vector expression.Knowledge vector is introduced into the part of word alignment and attention mechanism in text entailment recognition model.The experimental results show that the proposed method is better than the classical model.

关 键 词：文本蕴含 知识向量 深度学习 词对齐 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]