基于动态语义编码双向LSTM的中文语义相似度计算  

CHINESE SEMANTIC SIMILARITY CALCULATION BASED ON DYNAMIC SEMANTIC CODING BI-DIRECTIONAL LSTM

作　　者：初雅莉 郑虹[1] 侯秀萍[1] Chu Yali;Zheng Hong;Hou Xiuping(School of Computer Science and Engineering,Changchun University of Technology,Changchun 130000,Jilin,China)

机构地区：[1]长春工业大学计算机科学与工程学院,吉林长春130000

出　　处：《计算机应用与软件》2020年第6期224-229,共6页Computer Applications and Software

基　　金：吉林省教育厅项目(2015121,2015122)。

摘　　要：中文句子结构有较高的复杂性,导致计算语义相似度非常困难,准确率不高。针对这种情况,改进一种动态语义编码双向LSTM中文语义相似度计算模型。采用双向LSTM编码器解码器结构,提取由标准句子和自然句子组成的句子对的双向语义特征。在解码器结构中加入动态语义编码规则,调整输入到解码器的语义编码信息,并把编码信息存储到LSTM神经元节点中,从而得到句子对更准确的相似度值。模型在汉字数据集、混合数据集和原始数据集三种中文句子对数据集中进行实验。实验结果表明,该模型在三种数据集中语义相似度的计算优于其他循环神经网络模型。Due to the complexity of Chinese sentence structure,it is very difficult to calculate the semantic similarity,and the accuracy is not high.Aiming at this situation,we improve a bi-directional LSTM Chinese semantic similarity calculation model by dynamic semantic coding.It adopted a bi-directional LSTM encoder-decoder structure to extract bi-directional semantic features of sentence pairs consisting of standard sentences and natural sentences.A dynamic semantic coding rule was added to the decoder structure,the semantic coding information input to the decoder was adjusted,and the coded information was stored in the LSTM cell,so as to get more accurate similarity value of the sentence pair.The model was used to test three Chinese sentences in Chinese character dataset,mixed dataset,and original dataset.The experimental results show that the model is superior to other recurrent neural network models in the calculation of semantic similarity in the three datasets.

关 键 词：中文语义 语义相似度 双向LSTM 动态语义编码规则 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]