多重注意力特征融合网络对中文评价情感分析  

Sentiment Analysis of Chinese Evaluate Text Based on Multiple Attention and Feature Fusion Network Model

作　　者：王勇[1] 张索宇 吕心怡 WANG Yong;ZHANG Suo-yu;LV Xin-yi(School of Artificial Intelligence,Chongqing University of Technology,Chongqing 401135,China;College of Computer Science and Engineering,Chongqing University of Technology,Chongqing 400054,China)

机构地区：[1]重庆理工大学人工智能学院,重庆401135 [2]重庆理工大学计算机科学与工程学院,重庆400054

出　　处：《小型微型计算机系统》2021年第8期1633-1638,共6页Journal of Chinese Computer Systems

基　　金：国家社科基金西部项目(17XXW005)资助。

摘　　要：针对目前中文评价情感分析对深层情感语义信息关注较少的问题,提出一种多重注意力的特征融合神经网络模型简称MTA-CBG(Multi-Attention Convolution-BiGRU).传统词向量不能有效解决一词多义的情况,本文构建了自注意力(Self-Attention)词向量矩阵模型,获取词语间的关联特征.通过多尺度宽卷积结构(Multi-scale Wide Convolution,MWC)全面地提取局部特征.将两种不同粒度的特征融合后输入双向门限循环单元(Bidirectional Gated Reccurrent Unit,BiGRU)学习序列化特征,在解决长距离依赖问题的同时获取更广泛的文本特征.最后输入改进的高速注意力层(Attention-Highway)构建句子级的关联,提取深层情感语义特征.通过多组对比实验证明本文所提方法能有效提高中文评价情感分析的准确率和F1值.Aiming at the situation that the sentiment analysis of Chinese evaluate text pays less attention to deep emotional semantics,this paper presents a multi-attention and feature fusion network model M TA-CBG. Traditional w ord vector cannot solve the problem of polysemy,this paper builds a self-attention word vector matrix model to get related features between words. To extract the local features comprehensively by the multi-scale wide convolution structure(MWC). The two types of features are fused and input into bidirectional gated recurrent unit(Bi GRU) to learn serialized features. To acquire wider emotional semantic information while solving the problem of long-distance dependence. Finally,the features are input into the improved Attention-Highway layer to construct sentencelevel association,and to extract deep emotional features. Based on comparative experiments of the datasets,the results confirm the means in this paper can exactly enhance the accuracy and F1 score on sentiment analysis of Chinese evaluate text.

关 键 词：多重注意力 特征融合 多尺度宽卷积 双向门限循环单元 高速注意力层 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]