基于双向时间深度卷积网络的中文文本情感分类  

CHINESE TEXT SENTIMENT CLASSIFICATION BASED ON BIDIRECTIONAL TEMPORAL DEEP CONVOLUTIONAL NETWORK

作　　者：韩建胜 陈杰 陈鹏 刘杰 彭德中 Han Jiansheng;Chen Jie;Chen Peng;Liu Jie;Peng Dezhong(College of Computer Science(College of Software),Sichuan University,Chengdu 610065,Sichuan,China)

机构地区：[1]四川大学计算机学院(软件学院)

出　　处：《计算机应用与软件》2019年第12期225-231,共7页Computer Applications and Software

基　　金：国家自然科学基金青年基金项目(61602329);四川省科技基础条件平台项目(2018TJPT0031);川大-泸州战略合作项目(2017CDLZ-G25);四川大学—自贡市校地科技合作专项资金项目(2018CDZG-10)

摘　　要：普通时间卷积网络对文本进行单向特征提取不能充分捕捉文本特征,对文本的分析能力较弱。提出一种基于双向时间卷积网络(Bi TCN)的情感分析模型。模型使用单向多层空洞因果卷积结构分别对文本进行前向和后向特征提取,将两个方向的序列特征融合后进行情感分类。研究并分析模型中卷积层数、卷积核大小和空洞因子三个参数对情感分类结果的影响。实验证明,与单向时间卷积网络情感分析模型相比,双向时间卷积网络模型在四个中文情感分析数据集上的准确率分别提高了2.5%、0.25%、2.33%和2.5%。The unidirectional feature extraction of text with common temporal convolutional network cannot fully capture information of text,and it is weak in text analysis.Therefore,we propose a sentiment analysis model based on the bidirectional temporal convolutional network(Bi-TCN).The model used a one-way multi-layer dilated causal convolution structure to extract the forward and backward features of the text,and combined the sequence features in two directions for sentiment analysis.We studied and analyzed the influence of three parameters on sentiment classification results,that was convolution layer,convolution kernel size and dilation factor.Experiments show that compared with the unidirectional temporal convolutional network sentiment analysis model,the proposed model improves the Chinese sentiment analysis on four datasets by 2.5%,0.25%,2.33%,and 2.5%respectively.

关 键 词：情感分析 自然语言处理 空洞卷积 因果卷积 双向时间卷积网络 

分 类 号：TP3[自动化与计算机技术—计算机科学与技术]