Cheng Yanfen, Yuan Minghui, He Fan, Shao Xun, Wu Jiajun
School of Computer Science and Artificial Intelligence, Wuhan University of Technology, Wuhan, 430070, China.
Department of Electrical and Electronic Information Engineering, Toyohashi University of Technology, Japan.
Neural Netw. 2025 Jul 12;192:107864. doi: 10.1016/j.neunet.2025.107864.
Aspect sentiment classification (ASC) is a subtask of Aspect-based sentiment analysis (ABSA), and its goal is to predict the sentiment polarity corresponding to specific aspect word within a sentence. Existing ABSA methods have achieved impressive predictive performance by not only leveraging attention mechanism to capture semantic association information within sentences, but also employing graph convolutional network (GCN) to exploit the syntactic structure information of dependency and constituency trees. However, these methods still have two main shortcomings: (1) they overlook the utilization of local sentiment features in the constituency tree, focusing only on global sentiment features; and (2) they fail to comprehensively utilize the syntactic information from both dependency and constituency trees. To address these issues, we propose a novel multi-granularity graph convolutional network (MGCN), which comprises three main components: an attention layer, a mask matrix layer, and a GCN layer. In the attention layer, we constructed semantic association matrices using an attention mechanism to explore the semantic associations between words and overall sentence semantics. In the mask matrix layer, we designed hierarchical rule-based multi-granularity constituency tree mask matrices (CM) to extract sentiment features from local to global levels within the constituency tree. Additionally, to obtain a more comprehensive syntactic features set, we fully fused the structural characteristics of the dependency and constituency trees to create multi-granularity fusion mask matrices (FM), which were further enhanced by the semantic association matrices. Finally, in the GCN layer, we performed convolution operations on the enhanced FM to strengthen the node representations. Experiments on the SemEval 2014 and Twitter datasets demonstrated effectiveness of MGCN.
方面情感分类（ASC）是基于方面的情感分析（ABSA）的一个子任务，其目标是预测句子中与特定方面词相对应的情感极性。现有的ABSA方法不仅通过利用注意力机制来捕捉句子中的语义关联信息，还通过采用图卷积网络（GCN）来利用依存树和成分树的句法结构信息，取得了令人印象深刻的预测性能。然而，这些方法仍存在两个主要缺点：（1）它们忽略了成分树中局部情感特征的利用，仅关注全局情感特征；（2）它们未能全面利用依存树和成分树的句法信息。为了解决这些问题，我们提出了一种新颖的多粒度图卷积网络（MGCN），它由三个主要部分组成：一个注意力层、一个掩码矩阵层和一个GCN层。在注意力层，我们使用注意力机制构建语义关联矩阵，以探索单词与整个句子语义之间的语义关联。在掩码矩阵层，我们设计了基于层次规则的多粒度成分树掩码矩阵（CM），以从成分树内的局部到全局级别提取情感特征。此外，为了获得更全面的句法特征集，我们充分融合了依存树和成分树的结构特征，以创建多粒度融合掩码矩阵（FM），并通过语义关联矩阵对其进行进一步增强。最后，在GCN层，我们对增强后的FM执行卷积操作，以强化节点表示。在SemEval 2014和Twitter数据集上的实验证明了MGCN的有效性。
