Cross-Linguistic Transfer for Low-Resource Sentiment Analysis: Adapting Multilingual BERT to Dialects and Minority Languages

In this study, we propose Dia-mBERT, a cross-language migration framework based on the multilingual pre-training model (mBERT), to address the challenge of resource scarcity in minority language/dialect sentiment analysis through three-layer design of linguistic feature injection (Dialect Lexicon Expansion and Sub-word Reconstruction), phased migration learning (Intermediate Pre-training + Adapter Lightweight Fine-tuning), and adversarial domain adaptation (Gradient Inversion Layer for reducing the dialect-standard language differences). The three-layer design enables efficient sentiment analysis in low-resource scenarios such as Arabic, Cantonese, and Maori. Experiments show that Dia-mBERT improves F1 by 12.3% on average over the baseline model, and maintains 15% performance advantage with 500 labelled samples, especially in morphologically rich languages (e.g., Arabic) and extremely low-resource indigenous languages (e.g., Maori). The study also constructs a multi-dialect sentiment dataset, reveals the influence of language distance on the migration effect, and calls for the establishment of a multi-dimensional evaluation standard for NLP in low-resource languages, which will provide a technological paradigm for the protection of linguistic diversity in the digital age.