基于混合注意力机制的中文文本蕴含识别方法  

Chinese textual entailment recognition method based on hybrid attention

作　　者：黄生斌 肖诗斌[1,2] 都云程 施水才 HUANG Shengbin;XIAO Shibin;DU Yuncheng;SHI Shuicai(Computer School,Beijing Information Science&Technology University,Beijing 100192,China;TRS Information Technology Co.,Ltd,Beijing 100101,China)

机构地区：[1]北京信息科技大学计算机学院,北京100101 [2]拓尔思信息技术股份有限公司,北京100101

出　　处：《北京信息科技大学学报（自然科学版）》2020年第3期89-93,98,共6页Journal of Beijing Information Science and Technology University

摘　　要：将注意力机制应用于中文文本蕴含识别研究,提出了一种基于混合注意力机制的中文文本蕴含识别方法。为使模型能捕获文本间更丰富的语义信息,利用混合注意力机制获取文本词语级别和句子级别的语义信息,将不同级别的语义信息进行融合后对文本蕴含关系进行识别,提高了模型识别中文蕴含关系的准确性。模型的输入阶段采用了BERT字向量,提高了模型的泛化能力。该方法在CCL2018(第十七届中国计算语言学大会)中文文本蕴含识别测评任务验证集上准确率达80.38%,实验结果表明了该方法的有效性。Applying the attention mechanism to the study of Chinese textual entailment recognition,a method of Chinese textual entailment recognition based on hybrid attention mechanism is proposed.In order to enable the model to capture richer semantic information of texts,a hybrid attention mechanism is used to obtain semantic information at the word level and sentence level of the text,and to merge the semantic information at different levels before recognizing the text entailment relationship,which improves the accuracy of the model in recognizing the Chinese entailment relationship.In addition,the BERT word vector is used in the input stage of the model,which improves the generalization ability.This method has an accuracy rate of 80.38%in the verification set of CCL2018 Chinese textual entailment recognition tasks,the experimental results show that this method is effective.

关 键 词：中文文本蕴含 深度学习 混合注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]