Abstract: Aspect-level sentiment analysis, as an important task in the field of affective computing, aims to identify the emotional tendencies about specific aspects in texts. In order to improve the performance in this task, a network model (ESMFGCN) that enhances the fusion of syntactic features and multi-feature graph convolution is proposed, using dependency trees to represent the grammatical structural relationships between words in sentences. Due to the simple use of dependencies tree method will cause irrelevant noise problems when modeling. The phrase structure tree is introduced and the phrase tree is converted into a hierarchical phrase matrix. The adjacency matrix and hierarchical phrase matrix constructed by the dependency tree are merged as the initial matrix of the graph convolution network, used to enhance syntactic information. In order to more precisely capture the association between aspect words and the entire sentence, an attention mechanism is introduced to establish a more refined association between the aspect word context and the entire sentence, and extract semantic information through a graph convolutional network. Finally, a fusion layer is designed to fuse semantic information and syntactic information to improve the accuracy and robustness of aspect-level sentiment analysis. This article designed comparative experiments, ablation experiments, and sensitivity analysis experiments on the Restaurant, Laptop, and Twitter datasets respectively. The experimental results show that compared with other research methods, the method in this paper has achieved significant performance improvement, proving the effectiveness and superiority of the model.

Key words: aspect-level sentiment analysis, syntactic features, attention mechanism, graph convolutional network

摘要： 方面级情感分析作为情感计算领域的重要任务，旨在识别文本中关于特定方面的情感倾向。为了提高在这一任务中的性能，提出了一种增强句法特征与多特征图卷积融合的网络模型（ESMFGCN），利用依赖树表示句子中单词之间的语法结构关系，由于单纯的使用依赖树方法在建模时会引发不相关的噪声问题，引入了短语结构树，并将短语树转化层级短语矩阵，并将由依赖树构造的邻接矩阵和层级短语矩阵合并作为图卷积网络的初始矩阵，用于增强句法信息。为了更精细地捕捉方面词与整个句子之间的关联，引入了注意力机制，对方面词上下文和整个句子建立更为精细的关联，并通过图卷积网络提取语义信息。最后设计融合层用于融合语义信息与句法信息，从而提高方面级情感分析的准确性和鲁棒性。本文在Restaurant、Laptop、Twitter数据集上分别设计对比实验、消融实验和敏感性分析实验，实验结果表明，相较于其他研究方法，本文的方法取得了显著的性能提升，证明了模型的有效性和优越性。

关键词: 方面级情感分析, 句法特征, 注意力机制, 图卷积网络