Dynamic Multi-Scale Attention for Fine-Grained Aspect-Based Sentiment Analysis

Abstract This research focuses on improving Aspect-Based Sentiment Analysis (ABSA) by addressing challenges related to multiple aspects and sentiment ambiguities within sentences. We propose the DAMSA-(Dynamic aspect-aware multi-scale attention) Transformer, a novel model that integrates three key components: Dynamic Aspect-Aware Attention (DAA), which adjusts attention based on the complexity of each aspect; Multi-Scale Feature Fusion (MSFF), which combines sentiment information from different linguistic levels; and Contrastive Aspect Learning (CAL), which helps distinguish between similar and conflicting sentiments more effectively. The model was evaluated on the SemEval-2014 Laptop and Restaurant datasets, showing consistent improvements over existing transformer-based methods with up to 4.45% increase in accuracy and 2.1% in macro-F1 score. Additional analyses, including attention visualizations and class-wise performance, demonstrate the model’s capability to capture sentiment nuances more accurately.