A Comprehensive Review of Transfer Learning-Based Sentiment Analysis Using BERT and Transformer Models

Sentiment analysis is a task in Natural Language Processing (NLP) that detects polarity, opinions, emotions, or attitudes in text data. Traditional systems based on lexicons and hand-crafted features are relatively straightforward, but their adaptability is constrained by vocabulary, context, domain differences, and linguistic variation. Deep learning introduced CNN, LSTM, BiLSTM, GRU, and attention-based representations, while transformer-based pre-trained language models significantly improved contextual transfer. Transfer learning reduces the cost of target-domain annotation by leveraging knowledge from a source domain or a pre-trained model for a target sentiment task. This review examines transfer-learning methods for sentiment analysis, focusing on machine learning, artificial intelligence, natural language processing, Hugging Face Transformers, BERT, and the fine-tuning of pre-trained language models. The literature is organized into transfer-learning taxonomy, model development, BERT-based modifications, hybrid architectures, datasets, evaluation approaches, limitations, and emerging parameter-efficient techniques. The analysis highlights research gaps in source-domain selection, negative-transfer detection, low-resource and code-mixed evaluation, computational efficiency, explainability, continual adaptation, and reproducibility. The review suggests that promising future systems should not focus solely on increasing model size, but should identify relevant knowledge, adapt effectively, and maintain robustness across diverse domains and languages.