Fine-Tuning DeBERTaV3 with Adaptive Training Strategies for Fine-Grained Sentiment Analysis

Artificial intelligence (AI) plays a pivotal role in modern society, especially in business intelligence, where it drives insights for customer feedback analysis, targeted recommendations, and strategic marketing. Within AI, natural language processing (NLP) enables machines to interpret and analyze human language, with sentiment analysis being a vital subset focused on understanding nuanced emotions in text. This research introduces Fine-tuned DeBERTaV3 with Adaptive Training Strategies (FiTDeBERTaV3-ATS), a model designed for fine-grained sentiment analysis that integrates DeBERTaV3 with an attention mechanism, cross-fold training, and multi-sample dropout. Evaluated on the fine-grained Stanford Sentiment Treebank (SST-5) dataset, the proposed model achieved an improved accuracy of 62.40%, surpassing existing baselines and highlighting DeBERTaV3's effectiveness for this task. These tailored strategies address the unique challenges of SST-5, such as figurative language and a limited dataset size, significantly enhancing the model's ability to capture subtle sentiment distinctions.