Abstract: Aspect-based sentiment analysis is an important task in natural language processing. Existing methods generally adopt single-type nodes to model semantic associations between aspects and opinion words, failing to effectively distinguish the contributions of different node types to sentiment representation. Moreover, current models usually construct syntactic trees using dependency parsers for semantic analysis, ignoring fine-grained semantic features and implicit sentiment modification relations contained within words. To address these issues, a heterogeneous multi-source graph convolutional network (HMS-GCN) is proposed. First, a heterogeneous graph with differentiated constraints is constructed to capture semantic and sentiment information from multi-type nodes and their corresponding edges, utilizing a dual-channel attention mechanism to weigh node types and individual node importance. Second, sememe knowledge enhances semantic representation of initial textual features, and phrase-level structural information is integrated into the syntactic adjacency matrix to improve syntactic dependency comprehension. Finally, heterogeneous graph representations and enhanced knowledge features are combined for sentiment prediction. Experimental results on public datasets (Twitter, Lap14, Rest15, Rest16) demonstrate that the proposed method outperforms the baseline model ISSK-GCN, with accuracy improvements of 1.18, 1.33, 1.55, 2.41, and macro-F1 improvements of 2.16, 2.40, 0.69, 5.96, validating its effectiveness.

Key words: Keywords: aspect-based sentiment analysis, heterogeneous graph, dual-channel attention mechanism, graph convolutional network, sememe knowledge

摘要： 方面级情感分析是自然语言处理领域中的重要研究任务，现有方法普遍采用单一类型节点建模句子中方面词与意见词的语义关联，难以有效区分不同类型节点对情感表达的差异性贡献。此外，模型通常利用句法依赖解析器构建句法树进行语义分析，忽视了单词本身所蕴含的细粒度语义信息及文本中隐含的情感强度修饰关系。针对上述问题，提出一种异构增强和多源知识融合的图卷积网络模型（HMS-GCN）。首先，构建基于差异化约束的异构图网络，利用异构图中多类型节点及其相应边关系，多维度捕捉句子语义情感信息，并通过双通道注意力机制关注不同类型节点及节点内部的重要程度。其次，利用义原知识对初始文本特征进行语义增强，并在句法邻接矩阵的基础上融合层级短语结构信息，提升模型对句法依存关系的理解能力。最后，融合异构图网络信息和知识增强特征以进行情感分类预测。在公开数据集Twitter、Lap14、Rest15和Rest16上的实验结果表明，相较次优模型ISSK-GCN，所提方法在准确率上分别提升了1.18、1.33、1.55、2.41个百分点，在宏F1值上分别提升了2.16、2.4、0.69、5.96个百分点，验证了其有效性。

关键词: 方面级情感分析, 异构图, 双通道注意力机制, 图卷积网络, 义原知识