结合MacBERT与多层次特征协同网络的音乐社交评论情感分析模型  

Sentiment analysis model of music social commentary fused with MacBERT and multi⁃level feature collaborative network

作　　者：兰庆炜[1] 樊宁 LAN Qingwei;FAN Ning(Luxun School of the Arts of Yan’An University,Yan’an 716000,China;Henan Province Bureau of Statistics,Zhengzhou 450018,China)

机构地区：[1]延安大学鲁迅艺术学院,陕西延安716000 [2]河南省统计局,河南郑州450018

出　　处：《电子设计工程》2023年第7期36-41,共6页Electronic Design Engineering

基　　金：陕西省教育厅重点研究项目(19JZ061)。

摘　　要：为更好地解决传统模型特征捕捉能力不足,词向量语义表示不准确等问题,提出了结合MacBERT与多层次特征协同网络的音乐社交评论情感分析模型MacBERT-MFCN(MacBERT and Multi-level Feature Collaborative Network)。采用MacBERT模型提取评论文本特征向量,解决静态词向量无法表示多义词的问题;多层次特征协同网络结合双向内置注意力简单循环单元(Bidirectional Built in Attention Simple Recurrent Unit,BiBASRU)和多层次卷积神经网络(Multilevel Convolutional Neural Network,MCNN)模块,全面捕捉局部和上下文语义特征;软注意力用来衡量分类特征贡献的大小,赋予关键特征更高权重。基于网易云评论文本数据集进行实验,结果表明,MacBERTMFCN模型F1值高达95.56%,能有效地提升文本情感分类准确率。In order to solve the problems of insufficient feature capture ability of traditional models and inaccurate semantic representation of word vectors,a music social comment sentiment analysis model combining MacBERT and multi⁃level feature collaborative network is proposed.MacBERT model is used to extract the feature vector of comment text to solve the problem that the static word vector can not represent polysemy;The multi⁃level feature collaborative network combines the Bidirectional Build in Attention Simple Recurrent Unit(BiBASRU)and Multi⁃level Convolutional Neural Network(MCNN)modules to comprehensively capture local and contextual semantic features;Soft attention measures the contribution of classification features and gives higher weight to key features.Experiments based on the NetEase cloud comment text data set show that the F1 value of MacBERT-MFCN model is as high as 95.56%,which can effectively improve the accuracy of sentiment classification.

关 键 词：情感分析 MacBERT 多层次特征协同网络 SRU 软注意力 

分 类 号：TN-9[电子电信]