基于多层次知识增强的方面级情感分析模型  

Aspect-based sentiment analysis model based on multilevel knowledge enhancement

作　　者：段文杰 邓金科 张顺香[1,2,3] 李书羽 周若彤 DUAN Wenjie;DENG Jinke;ZHANG Shunxiang;LI Shuyu;ZHOU Ruotong(School of Computer Science and Engineering,Anhui University of Science and Technology,Huainan 232001,China;Artificial Intelligence Research Institute,Hefei Comprehensive National Science Center,Hefei 230088,China;School of Computer,Huainan Normal University,Huainan 232038,China)

机构地区：[1]安徽理工大学计算机科学与工程学院,安徽淮南232001 [2]合肥综合性国家科学中心人工智能研究院,安徽合肥230088 [3]淮南师范学院计算机学院,安徽淮南232038

出　　处：《智能系统学报》2024年第5期1287-1297,共11页CAAI Transactions on Intelligent Systems

基　　金：国家自然科学基金面上项目(62076006);认知智能全国重点实验室开放课题(COGOS-2023HE02);安徽高校协同创新项目(GXXT-2021-008).

摘　　要：在方面级情感分析任务中,现有研究侧重于挖掘评论语句的语义信息和句法依赖约束,未能综合考虑情感知识、概念知识和单词之间的句法依赖类型对方面情感倾向判别准确性的影响。针对这一问题,提出一种基于多层次知识增强的方面级情感分析模型(multilevel knowledge enhancement,MLKE),利用外部知识对评论语句进行情感、句法和概念3个层次的知识增强。首先,利用情感知识及单词之间的依赖类型来增强句子的依赖图,并通过图卷积网络建模节点特征,得到情感和句法增强的特定方面表征;其次,利用概念图谱对方面词概念增强后,与特定方面表征进行融合,得到多层次知识增强的方面表征;最后,采用交互注意力机制实现上下文表征与方面表征之间的协调优化。5个公共数据集上的实验结果表明,所提模型的准确率和宏F1值均得到提高。In aspect-based sentiment analysis tasks,mining semantic information and syntactic dependency constraints from comment sentences is a key focus in existing research.However,this often underestimates the influence of comprehensive factors,including sentiment knowledge,conceptual knowledge,and syntactic dependency types between words on aspect sentiment orientation judgment.To address this problem,we propose an aspect-based sentiment analysis model based on multilevel knowledge enhancement(MLKE),which uses external knowledge to enhance the knowledge of comment sentences on three levels:sentiment,syntax,and concept.First,sentiment knowledge and the dependency types between words are employed to enhance the dependency graph of sentences.Specific aspect representations containing sentiment and syntactic enhancements are obtained through graph convolution networks that model modular node features.Second,to obtain multilevel knowledge-enhanced aspect representation,the concept graph is used to enhance the conceptual understanding of aspect words,and then the aspect word representation is fused with the specific aspect representation obtained in the previous step.Finally,the coordination and optimization between context representation and aspect representation is achieved using an interactive attention mechanism.The experimental results show that the accuracy and macro-F1 values of the model are improved on five datasets.

关 键 词：方面级情感分析 知识增强 情感知识 句法知识 概念知识 依赖图 图卷积网络 交互注意力机制 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]