A hybrid RoBERTa–BiLSTM framework for aspect-based sentiment analysis of Monkeypox tweets

The Monkeypox and COVID-19 outbreaks have brought into focus the significance of social media as a real-time source of public health information, where people share their opinions and concerns regarding different health issues. However, the task of reliable aspect-level sentiment extraction is difficult due to the presence of ambiguous language and class imbalance. This paper presents a hybrid model that uses RoBERTa and stacked BiLSTM layers for aspect-based sentiment analysis (ABSA). The proposed model utilizes the capability of transformers to represent context and the ability of BiLSTM to learn sequential dependencies for fine-grained sentiment classification. The proposed approach is rigorously tested on the primary Monkeypox X (Twitter) large-scale dataset, a generalization COVID-19 Twitter corpus, and the SemEval-2014 ABSA benchmark using accuracy, precision, recall, F1-score, five-fold cross-validation, and one-way ANOVA testing. The experimental results show that the proposed approach performs better than traditional machine learning, deep learning, and transformer-based baselines, achieving an accuracy of 95.17% and an F1-score of 95.04% on the Monkeypox dataset, 92.15% accuracy and 91.87% F1-score on the COVID-19 dataset, and 87.68% accuracy and 87.12% F1-score on the SemEval-2014 dataset. Statistical analysis proves the significance of the performance improvements across all evaluated domains.