A Comparative Study of Attention Mechanisms in Deep Learning Models for Aspect-based Sentiment Analysis of Customer Reviews

Aspect-Based Sentiment Analysis (ABSA) refines conventional sentiment analysis by targeting specific aspects within customer reviews. However, existing deep learning approaches often fall short when multiple aspects such as price, quality, or service are discussed in a single review, leading to misclassifications. In this paper, we present a comparative study of four baseline models (Recurrent Neural Network, Convolutional Neural Network, Recursive Neural Network, and Memory Network) alongside four attention mechanisms (Self, Multihead, Global, and Hierarchical). Experiments on publicly available datasets reveal that while the RNN baseline achieves the best accuracy (73.1%) among non-attention models, incorporating attention substantially enhances performance. In particular, RNN with Hierarchical Attention attains the highest accuracy of 84.7%, highlighting its superiority in capturing both local (word-level) and global (sentence-level) dependencies. Global, Self and Multihead Attention boosts performance, but offer moderate gains. The study shows that adding attention mechanisms greatly improves multi-aspect sentiment classification. Self-Attention increased performance by an average of 2.5%, Multi-Head Attention by 6.8%, Global Attention by 10.2%, and Hierarchical Attention by 15.9%, making it the most effective. These findings underscore the importance of attention mechanisms for fine-grained sentiment tasks and inform researchers on selecting the most effective attention strategy for ABSA. We conclude by suggesting avenues for future work, including aspect-specific attention enhancements to further refine model accuracy.