A Self-Attention-Based Multi-Level Fusion Network for Aspect Category Sentiment Analysis

Aspect category-based sentiment analysis (ACSA) is a fine-grained sentiment analysis approach for predicting the sentiment polarity associated with the different aspect categories in a text. While numerous strategies have incorporated attention mechanisms to generate contextual features for specific aspectual objectives, their sole use may be influenced by the sentiments conveyed in other aspect categories. This can occur when aspectual targets are absent from the sentence and when words carry varying contextual sentiments across multiple aspect categories. This paper introduces the attention-based multi-level feature aggregation (AMFA) network, which simultaneously considers local and global information by applying attention to convolutional filters. It further employs context-guided self-attention modules at multiple levels to efficiently amalgamate learned different aspects and the interactions between contextual features within a cohesive framework. We tested the proposed approach on four public datasets. The results demonstrated the efficiency of our model in extracting more precise semantics and sentiments related to specific feature categories. To examine the practicality of AMFA, we constructed an online review dataset of table grapes from an e-commerce platform, where our model outperformed the baseline models with an accuracy of 88% and a macro-averaged F-1 score of 73.23%. We also validated the effectiveness of each AMFA module by testing them separately on all datasets. Experimental results prove that our proposed model is adept at semantically separating aspect embeddings from words in sentences and minimizing the impact of irrelevant information. Additionally, the robustness of our model is substantiated by supplementary experiments conducted on a constructed Chinese dataset.