基于BERT的孪生网络计算句子语义相似度  

Siamese Network Computing Sentence Semantic Similarity Based on BERT

作　　者：李景玉 LI Jingyu(Beijing Polytechnic,Beijing,100176 China)

机构地区：[1]北京电子科技职业学院,北京100176

出　　处：《科技资讯》2021年第32期1-4,共4页Science & Technology Information

基　　金：2021年北京电子科技职业学院院内科技类课题(项目编号:2021Z030-KXY)。

摘　　要：孪生网络因自身特点被应用于句子相似度计算,但若模型只关注构成句子本身的词语,是无法获得该词语更多的语义信息的。针对这一问题,该文基于BERT模型,提出基于BERT的孪生网络计算句子语义相似度的方法,该方法通过BERT模型,得到具有丰富语义信息的句子向量,再利用该向量在孪生网络中计算相似度。在公开数据集LCQMC上的实验结果表明,基于BERT的孪生网络计算句子语义相似度方法的Accuracy优于基于卷积孪生网络计算句子语义相似度方法。Siamese Network is applied to sentence similarity computation due to its own characteristics.However,if the model only focuses on the words that constitute the sentence itself,it can not obtain more semantic information of the words.In order to solve this problem,this paper proposes a method to calculate sentence semantic similarity based on Siamese Network based on BERT model.Through BERT model,the sentence vector with rich semantic information is obtained,and then the vector is used to calculate the similarity in Siamese Network.The experimen‐tal results on the public data set LCQMC show that the Accuracy of the Siamese Network method based on BERT is better than that based on convolution Siamese Network.

关 键 词：BERT 孪生网络 句子向量 句子相似度 

分 类 号：TP312[自动化与计算机技术—计算机软件与理论]