Dataset distribution.

Aspect-level sentiment analysis is a significant task in the field of natural language processing. It can process text in a fine-grained manner to predict the sentiment polarity of a specific aspect word in a sentence. However, existing single-channel models often ignore high-dimensional local feature information in syntactic dependencies, have a single structure, and cannot fully extract text features. At the same time, there are often multiple opinion words with diverse sentiment attitudes in a sentence, so there is a certain amount of noise when processing features, which interferes with the model’s understanding of the sentiment semantics related to aspect terms. To address the problems, this paper proposes an aspect-level sentiment analysis model (MSDC) based on multi-scale dual-channel feature fusion. First, through multi-head gated self-attention channels and graph neural network channels, the model further enhances its understanding of the spatial hierarchical structure of text data and improves the expressiveness of features. Then, we design an adaptive feature fusion mechanism that dynamically adjusts the weight ratio of aspect words to context according to a given aspect. Hence, the task pays more attention to key information. Finally, the data is integrated and processed through a capsule network. The results indicate that our model exhibits superior effectiveness on multiple public datasets, especially when processing fine-grained text sentiment analysis tasks, significantly improving the accuracy and F1 value compared to existing technologies.

方面级情感分析（Aspect-level Sentiment Analysis）是自然语言处理（Natural Language Processing）领域的一项重要任务，其可通过细粒度方式处理文本，以预测句子中特定方面词的情感极性。然而，现有单通道模型往往忽略句法依存关系中的高维局部特征信息，且结构单一，无法充分提取文本特征。同时，单句中常存在多个情感倾向各异的评价词，因此在处理特征时会引入一定噪声，干扰模型对与方面词相关的情感语义的理解。针对上述问题，本文提出一种基于多尺度双通道特征融合的方面级情感分析模型（MSDC）。首先，模型通过多头门控自注意力通道与图神经网络（Graph Neural Network）通道，进一步增强对文本数据空间层级结构的理解，提升特征表达能力；随后，设计自适应特征融合机制，可根据给定方面动态调整方面词与上下文的权重占比，使任务更聚焦关键信息；最后，通过胶囊网络（Capsule Network）完成数据整合与处理。实验结果表明，所提模型在多个公开数据集上展现出更优的有效性，尤其在处理细粒度文本情感分析任务时，相较于现有技术，其准确率与F1值均得到显著提升。