Transfer Learning for Cross-Domain Sentiment Classification via Domain Adaptation

The exponential proliferation of user-generated content across online platforms has driven a substantial surge in domain-specific textual data. Consequently, sentiment analysis has emerged to facilitate context-aware, real-time decision making in dynamic, data-intensive domains. Nevertheless, obtaining large-scale annotated data for all domains remains impractical, therefore a necessity for robust cross-domain sentiment classification techniques is still necessary. This work proposes WBiGe a hybrid domain adaptation architecture that combines static and contextual embeddings with a hierarchical BiGRU encoder to exploit semantic structure globally and sentiment cues locally. A lightweight convolutional architecture is used to extract phrase level patterns that are usually ignored by sequential encoders to improve short range sentiment expressions detection. To focus sentiment informative pivots and nonpivots during domain drilldown, a class aware domain alignment mechanism based on weighted maximum mean discrepancy is utilized. Furthermore, an iterative pseudolabeling strategy and dual-level attention modules are incorporated into the architecture to continuously improve target domain prediction to increase the interpretability of the model. The proposed WBiGe framework was extensively assessed using the Amazon benchmark dataset, in which it significantly outperforms the existing models across a wide range of cross-domain classification tasks with a record peak accuracy of 92.75%.