Aspect-based Sentiment Classification (ABSC) has become a vital task in natural language processing that enables fine-grained sentiment detection at the aspect level rather than general polarity analysis. Still, existing methods have issues in capturing deep semantic context for preserving hierarchical dependencies and efficiently tuning hyperparameters, which leads to suboptimal performance in noisy and real-world textual data. To address the limitations, this research proposes a hybrid deep learning model that integrates RoBERTa-based contextual embeddings with the Transformer encoder and Capsule Networks (CapsNets) for enriched semantic and structural feature extraction. Here, aspect-relevant words are highlighted by the self-attention mechanism to refine the features, and the Improved Starfish Optimization Algorithm (ISFOA) is presented for hyperparameter optimization using chaotic sinusoidal mapping. Experiments on the publicly accessible dataset show that the proposed approach outperforms current approaches in terms of accuracy, F1-score, MCC, and other important metrics.
Nagelli et al. (Sat,) studied this question.