Transformer-Driven Machine-Learning Framework for Low-Resource Sentiment Classification

Sentiment classification in under-resourced languages remains challenging because of scarce annotated corpora, linguistic complexity, and computational constraints. Indian regional language Malayalam further intensifies these challenges through its agglutinative morphology and limited labeled resources. This study systematically evaluates a Transformer-assisted machine-learning framework for Malayalam sentiment analysis. Across two datasets, five multilingual Transformer encoders were paired with eight classical classifiers, yielding 80 Transformer–ML configurations and 10 end-to-end baselines. Each Transformer was first fine-tuned using a task-specific neural head; its task-adapted encoder was then frozen to generate attention-mask-aware mean-pooled sentence embeddings for classical classification. The same checkpoint was used for end-to-end evaluation and embedding extraction, permitting comparison between neural and classical classifiers. Experiments used a monolingual Malayalam Sports News dataset and the Dravidian CodeMix dataset. Performance was evaluated using standard classification metrics and reported as mean ± standard deviation across multiple random seeds. Robustness was further assessed using independent stratified data splits for the selected best-performing configurations to evaluate sensitivity to data-partition variations. GPU-based embedding–extraction and CPU-based classifier-prediction latencies were measured separately. Paired performance differences across seeds were assessed using the Wilcoxon signed-rank test. The Transformer–ML configurations achieved comparable classification performance relative to their end-to-end counterparts, although inference-time characteristics varied with the downstream classifier. These findings provide practical guidance for selecting suitable Transformer–classifier combinations for diverse linguistic scenarios.