Two-Stage Bengali Sentiment Classification: Domain Adaptation Through Continual Learning and Parameter-Efficient Fine-Tuning
This work addresses the challenge of sentiment classification for low-resource languages like Bengali under domain data scarcity by proposing SentiBanglaBERT, a novel framework employing a two-stage strategy. First, it enhances contextual adaptability through continued pretraining on news-domain corpora; second, it leverages Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning. The approach innovatively integrates domain adaptation with interpretability analysis, utilizing SHAP to uncover the influence of key Bengali morphological features—such as negation suffixes and aspect markers—on sentiment predictions. Experimental results demonstrate that the model achieves performance comparable to strong baselines while maintaining computational efficiency and offering linguistically insightful explanations.