Multi-granularity alignment and discriminative enhancement based multi-source cross-domain text classification

Multi-Source Cross-Domain Text Classification (MSCDTC) aims to improve the performance of the target domain by utilizing knowledge from multiple source domains, but existing methods face two major bottlenecks. Firstly, the current model faces the prominent problem of single alignment strategy limitation in domain alignment. Secondly, classifiers struggle to fully learn strong discriminative features between categories in complex distributions from multiple sources. To address the above issues, we propose the Multi-Granularity Alignment and Discriminative Enhancement Based (MGADE) method. At first, the Mean Maximum Mean Discrepancy (MMMD) and Mean Local Maximum Mean Discrepancy (MLMMD) are proposed to establish a complementary dual-grained alignment mechanism that adaptively fuses global domain consistency with local substructure matching, thereby overcoming the limitations of single-grained alignment strategies and enabling more accurate and robust cross-domain knowledge transfer. In addition, to enhance inter-class discriminability and enforce robust decision boundaries under multi-source distributional shifts, the Adaptive Adversarial Multi-Classifier Fusion Mechanism including a main classifier and a Gradient-Driven Adversarial Bias-Enhanced Classifier (GABC) is constructed. What is more, considering that labels have specific meanings in text classification and that the cosine similarity between labels and texts can highlight words that have a greater impact on text classification, different label-based embedding vector update strategies for the source and target domains are respectively proposed. Experimental results from multiple comparison experiments on sentiment text classification and news text classification datasets demonstrate that the model proposed in this paper outperforms the state-of-the-art method. The main code is available at “https://github.com/sgysgywaityou/MGADE”.