Few-shot adaptation improves minority-label Bangla hate speech detection

A two-phase machine-learning framework called FewShot AdaptiveMLC improved classification performance on an imbalanced, multi-label Bangla hate speech dataset, according to a study in the International Journal of Data Science and Analytics. The researchers reported an absolute increase of 0.072 in macro-F1, from 0.712 to 0.784, compared with standard multi-task learning and fine-tuning baselines. The results come from the study’s experiments with a manually annotated Bangla dataset and do not establish how the method would perform in a deployed content-moderation system.

The framework addresses a specific problem in multi-label classification, where examples can belong to several hate-speech categories at once and some categories have substantially fewer training examples than others. In the first phase, multi-task learning uses a shared XLM-RoBERTa encoder to build a representation informed by multiple labels. Task-specific heads make predictions for the different categories while the encoder provides a common foundation.

The study reports that conventional oversampling and data augmentation did not resolve the weak performance on minority labels, motivating a method that adapts underrepresented categories separately rather than treating every label in the same way. FewShot AdaptiveMLC adds a second, targeted stage of few-shot fine-tuning for minority labels. During this phase, the shared encoder is frozen, and only the task-specific heads are updated using small, balanced few-shot subsets.

The approach therefore separates shared representation learning from label-specific specialization while preserving the encoder learned in the first phase. The authors describe this gradient-isolated arrangement as enabling focused adaptation without catastrophic forgetting. In the reported results, performance for underrepresented labels improved while performance for majority classes remained stable, supporting the use of a dedicated adaptation stage for the labels that continued to perform poorly after multi-task training.

Evaluation used BOISHOMMO, a new manually annotated multi-label Bangla hate speech dataset introduced as part of the research. The paper reports that FewShot AdaptiveMLC outperformed standard multi-task learning and fine-tuning baselines. Its macro-F1 increased to 0.784 from 0.712, while the Disability label reached an F1 score of 0.689 compared with 0.506 previously.

The study presents this design as a practical and scalable approach for multi-label hate speech detection in low-resource natural language processing settings, although the supplied research record does not provide the dataset’s full size, annotation procedure, train-test split, statistical uncertainty, or complete label-by-label results. The findings are limited to the dataset, language, and model configuration evaluated in the study. Because the available evidence is an abstract and article preview, it does not show whether the reported gains transfer to other hate-speech datasets, languages, encoder models, or operational moderation settings.

It also does not establish the number of examples used in each few-shot subset or provide uncertainty estimates for the reported differences. Within those limits, the study’s clearest measured result is that sequential training with a frozen shared encoder and targeted head updates produced higher macro-F1 and a substantially higher F1 score for the Disability label on BOISHOMMO, while maintaining stable performance for majority classes.

Subject of Research: Few-shot adaptation for imbalanced multi-label Bangla hate speech classification

Article Title: FewShot AdaptiveMLC: sequential multi-task learning approach with targeted few-shot adaptation to address class imbalance in multi-label datasets

Article References: Dip, S. D., Kafi, M. A. A., Moni, R., & Banshal, S. K. (2026). FewShot AdaptiveMLC: sequential multi-task learning approach with targeted few-shot adaptation to address class imbalance in multi-label datasets. International Journal of Data Science and Analytics, 22(1), Article 349. https://doi.org/10.1007/s41060-026-01306-3

Image Credits: AI Generated

DOI: 10.1007/s41060-026-01306-3

Keywords: few-shot learning, multi-task learning, class imbalance, Bangla hate speech, multi-label classification, XLM-RoBERTa, natural language processing, machine learning, minority labels, FewShot, AdaptiveMLC, sequential