Adaptive Attention and Optimization-Based Feature Refinement for Transformer-Driven Sentiment Classification

Sentiment analysis has become a fundamental task in natural language processing to extract public opinions from social media and online review platforms at a large scale. Recent machine learning, deep learning and transformer-based approaches have been significantly improved for sentiment classification. However, these models are still suffering from high-dimensional feature representations, redundant contextual information and unstable feature learning, which limit the classification accuracy and model generalization. To solve these problems, this paper proposes an Optimization-Guided Feature Refinement and Adaptive Attention RoBERTa (OGFAA-RoBERTa). OGFAA-RoBERTa reduced the feature dimensionality by an optimization-based feature selection technique and improved the model stability by a deep contextual feature extraction strategy. The OGFAA-RoBERTa incorporates preprocessing of data, optimization-guided feature refinement, adaptive attention learning and enhanced RoBERTa classifier to learn highly discriminative semantic representation while suppressing redundant contextual features. Experimental evaluation was performed on Sentiment140 and SemEval-2014 datasets. The OGFAA-RoBERTa achieves 92.34% accuracy on Sentiment140 and 88.76%, 92.43% and 85.74% on the Laptop14, Restaurant14 and Twitter subsets of SemEval-2014, respectively. The results show that the optimization-guided feature refinement and the adaptive attention can significantly improve the performance of sentiment classification, model stability and cross-domain generalization over the existing state-of-the-art methods.