Adaptive BERT-CNN Model for Efficient Maintenance of Variable Length Reviews in Rating Prediction

Ratings are users’ feedback used for many purposes, especially for developing and evaluating the quality of a recommender system. Text reviews often contain detailed emotions that comprehensively capture the preferences of users. The ability to effectively handle and analyze long reviews is crucial for various applications, ranging from sentiment analysis to classifications and recommendations. Traditional models, including BERT, excel at understanding contextual information within limited token lengths. In this paper, we propose a methodology that combines the strengths of BERT with CNN through a novel chunking approach to generate a continuous polarity to overcome the limits of text length. The proposed model, namely Adaptive BERT-CNN, optimizes the input for BERT by segmenting long texts into smaller, manageable chunks while preserving the overall context. It also allows CNNs to capture diverse features across multiple chunks that enhance the model’s ability to realize the patterns and relationships within the text so that the long reviews could help to understand users’ sentiments towards items. In our study, we conducted a comprehensive comparison across multiple datasets using baseline models like BERT and CNN, evaluating performance based on metrics such as accuracy and MSE. Our proposed model demonstrated a superior ability to detect context in long texts, achieving the highest accuracy of 99.76% on one dataset, significantly outperforming the baseline models.