An Attention Enhanced CNN BERT Model with TF IDF Feature Fusion for Cross Domain Sentiment Analysis

The exponential growth of online reviews has made the development of sentiment analysis models capable of reliable cross-domain generalization increasingly critical. Although transformer-based models such as BERT offer state-of-the-art contextual understanding, they often perform domain adaptation tasks when used in isolation. In this paper, we propose a novel sentiment classification model that integrates TF-IDF to extract the most significant lexicon features and CNN-BERT to extract semantic features fused with an attention mechanism. The model is trained in Amazon product reviews and evaluated in movie and course reviews. Experiments demonstrate that the proposed approach outperforms the TF-IDF+ CNN model in both accuracy and domain transferability. This study underscores the effectiveness of multi feature fusion and attention-guided weighting to build more adaptable and interpretable sentiment analysis systems.