A Self-Attention-Based Roberta Model for Opinion Mining in Sentiment Analysis

Sentiment analysis is important for analysing opinions and emotions expressed in text, particularly in applications involving social media monitoring and customer feedback analysis. In this paper a novel methodology for opinion mining is introduced based on self-attention-based Roberta model for this purpose and what makes it unique is the prepossessing, feature extraction and feature selection advanced techniques applied in the research work. In particular, the recently proposed human-engineered preprocessing pipeline, the NSEFA (Normalized Stack Error Flush Algorithm) step, tackles data noise and seasonality. PAVA is used for feature extraction, and the SSSO is utilized for feature selection. We present the learned model that combines Bi-LSTM and self-attention, which can learn a rich representation of the input text that also performs very well on the IMDb movie review dataset. The approach is also designed to be scalable and adaptable for real-time applications in various natural language processing tasks. Experimental results achieved 99.8% accuracy, 99.4% precision & 99.7% recall, recommending the robustness and performance of the model in sentiment classification.