计算机科学与探索

• 学术研究 •    下一篇

融合动态边权重图注意力网络的方面级情感分析

仲兆满, 吕慧慧, 张渝, 崔心如, 樊继冬, 黄泽宇   

  1. 1. 江苏海洋大学计算机工程学院,江苏 连云港 222005
    2. 江苏省海洋资源开发研究院,江苏 连云港 222005

Aspect-Level Sentiment Analysis with Integrated Dynamic Edge-Weighted Graph Attention Network

ZHONG Zhaoman,  LYU Huihui,  ZHANG Yu,  CUI Xinru,  FAN Jidong,  HUANG Zeyu   

  1. 1.School of Computer Engineering, Jiangsu Ocean University, Lianyungang, Jiangsu 222005, China
    2.Jiangsu Institute of Marine Resources Development, Lianyungang, Jiangsu 222005, China

摘要: 在方面级情感分析领域,传统研究多依赖静态图神经网络来建模文本上下文与方面之间的依赖关系。然而,这些方法往往忽略了句法结构和词性信息的动态特性,限制了模型在捕捉复杂语义关系方面的能力。为此,提出了一种动态边权重图注意力网络(Dynamic Edge-Weighted Graph Attention Network, DEWGAT),旨在更全面地捕捉文本的语义、句法和方面信息。DEWGAT模型通过整合BERT嵌入、句法距离矩阵、方面相关掩码和语法权重矩阵,动态计算基于句法依存关系的边权重,从而强化关键情感依存关系的建模。在图注意力网络中,引入了一种新的注意力机制,该机制同时考虑了方面词和语法依存关系,显著增强了对关键情感信息的捕捉能力,提高了情感分类的准确性。在实验中,DEWGAT模型在四个公开基准数据集上进行了评估,结果显示其在Twitter、Lap14和MAMS数据集上的准确率分别提升了1.16%、0.15%和4.85%。特别是在MAMS数据集上,DEWGAT模型的准确率达到了0.8937,显著优于多数对比模型,展示了其在处理复杂情感分析任务中的优越性能。尽管DEWGAT模型在多个数据集上表现出色,但目前尚未实现端到端的框架集成。

关键词: 动态边权重, 图注意力网络, 语法权重矩阵, 方面级情感分析

Abstract: In the field of aspect-level sentiment analysis, existing research mainly uses static graph neural networks to model the dependency relationship between context and aspects, ignoring the dynamic nature of syntactic structure and part-of-speech information. A Dynamic Edge-Weighted Graph Attention Network (DEWGAT) is proposed for aspect-level sentiment analysis. The model integrates BERT embeddings, syntactic distance matrix, aspect-related mask, and grammatical weight matrix to comprehensively capture the semantic, syntactic, and aspect information of the text. Dynamically calculate edge weights based on syntactic dependencies, reinforcing key emotional dependency relationships. Introduce an attention mechanism that considers aspect terms and grammatical dependencies in the graph attention network, enhancing the capture of crucial sentiment information and improving classification accuracy.Experimental results on four public benchmark datasets show that DEWGAT improved accuracy by 1.16%, 0.15%, and 4.85% on Twitter, Lap14, and MAMS datasets respectively. Notably, on the MAMS dataset, the accuracy reached 0.8937, significantly outperforming most comparison models. Although the model performs well, it has not yet achieved an end-to-end framework.

Key words: dynamic edge weight, graph attention network, grammatical weight matrix, aspect-level sentiment analysis