融合句法增强与语义增强的方面情感分析  

Fusion of syntactic enhancement and semantic enhancement for aspect-based sentiment analysis

作　　者：刘耀 吴云飞 周红静 黄少年 张震 LIU Yao;WU Yunfei;ZHOU Hongjing;HUANG Shaonian;ZHANG Zhen(Changsha Social Laboratory of Artificial Intelligence,Hunan University of Technology and Business,Changsha 410205,China;School of Digital Media Engineering and Humanities,Hunan University of Technology and Business,Changsha 410205,China;School of Computer Science,Hunan University of Technology and Business,Changsha 410205,China)

机构地区：[1]湖南工商大学长沙人工智能社会实验室,湖南长沙410205 [2]湖南工商大学数字媒体工程与人文学院,湖南长沙410205 [3]湖南工商大学计算机学院,湖南长沙410205

出　　处：《智能科学与技术学报》2025年第1期54-63,共10页Chinese Journal of Intelligent Science and Technology

基　　金：长沙人工智能社会实验室资助项目;国家自然科学基金项目(No.62202161);湖南省教育厅科学研究项目(No.21A0376);湖南省社会科学成果评审委员会课题(No.XSP25YBC164)。

摘　　要：使用图神经网络在句子的句法依赖树上建模进行方面情感分析时,主要关注句子的句法结构,但是忽视了对依赖树本身以及语义信息的研究,同时非结构化文本中句法信息会带来额外噪声。针对上述问题,提出一种融合句法增强与语义增强的图卷积网络模型,使用注意力机制提取方面信息和全局信息并通过门控机制结合增强语义信息,同时利用外部情感知识以及单词的距离特征重塑依赖树来提高意见词权重增强句法信息,最后对句法信息和语义信息进行动态特征融合。在3组公开数据集上进行了大量的实验,结果表明,提出的模型准确率和宏F1值相较于对比模型更高,验证了融合句法增强和语义增强在方面情感分析中的有效性。The graph neural networks mainly focus on the syntactic structure when they are used to model the syntactic dependency tree of a sentence for aspect-based sentiment analysis(ABSA).However,these methods neglect the study of dependency tree itself and semantic information,and syntactic information in unstructured text can bring additional noise.To address these issues,a graph convolutional network(GCN)model integrating syntactic and semantic enhancements was proposed.Attention mechanism was used to extract aspect information and global information,and enhance semantic information through gating mechanism.Simultaneously,dependency tree was reshaped to increase the weight of opinion words and enhance syntactic information by utilizing external emotional knowledge and the distance features of words.Finally,dynamic feature fusion was performed on semantic and syntactic information.Extensive experiments were conducted on three public datasets.The experiment results show that the accuracy and macro-F1 values are better than the compared models,which indicate the effectiveness of the fusion of syntactic enhancement and semantic enhancement for ABSA.

关 键 词：方面情感分析 图卷积网络 注意力机制 门控机制 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]