Aspect-Level Sentiment Classification of Consumer Reviews Utilizing BERT and Category-Aware Multi-Head Attention

Abstract

In recent years, the explosive growth of user-generated review texts has underscored the academic and societal significance of sentiment analysis. Although deep learning has achieved remarkable progress in this field, existing aspect-based sentiment classification (ABSC) methods face challenges in capturing the dynamic nature of sentiment categories. Furthermore, these methods often lack explicit modeling of category information, limiting their ability to adapt attention distributions based on sentiment categories. To address these challenges, this paper proposes a BERT-based model with a category-aware multi-head attention mechanism. The model introduces an aspect projection layer that maps aspect word embeddings into a feature space aligned with the context, thereby enhancing interaction between aspect words and the surrounding text. Additionally, a category-aware multi-head attention mechanism combines category weights and dynamic content weights to effectively fuse sentiment category information. This design significantly improves the model’s ability to capture sentiment features of multiple categories. Experimental evaluations on SemEval public datasets demonstrate that the proposed method outperforms state-of-the-art techniques, and ablation studies further confirm the effectiveness of its design.