LoRA-Enhanced BERT with Contrastive Learning for Political Sentiment Analysis
Ramesh Babu Pittala, Shruti Agarwal, Medikonda Asha Kiran, Manyam Thaile, Muhammad Ahmad Khan, G.Sai Aparna · 2025
This paper presents a strong deep-learning model that aims to improve sentiment classification in political speech. The model combines BERT embeddings, adapted using Low-Rank Adaptation (LoRA), with contrastive learning based on SimCLR and a cross-attention fusion mechanism. This blend enables the model to successfully learn contextual subtleties prevalent in political texts, including sarcasm, ambiguity, and nuanced tone shifts. The model reads data from Reddit posts in the politics category and official parliamentary debates, providing rich, varied textual input. Experimental outcomes of largescale tests show exceptional performance, with $96 \%$ accuracy and a 93% F1 score, surpassing conventional baselines such as vanilla BERT and CLIP-based classifiers. In addition to enhanced accuracy, the architecture is scalable and maintains consistent performance across various policy environments. This method offers policymakers actionable insights into public opinion and debate dynamics by facilitating scalable sentiment analysis suitable for real time settings.