LLM-Based Multi-Task Bangla Hate Speech Detection: Type, Severity, and Target

TL;DR

This paper introduces BanglaMultiHate, the first large-scale multi-task hate speech dataset for Bangla that labels type, severity, and target. It systematically compares classical models, Bangla-specific pretrained models, and large language models under zero-shot prompting and LoRA-based fine-tuning, highlighting the critical role of language-specific pretraining and task-focused adaptation in low-resource settings. The findings show BanglaBERT often outperforms zero-shot LLMs and that LoRA fine-tuning narrows the gap with monolingual PLMs, though fully surpassing language-specific pretraining remains challenging. The dataset and accompanying scripts are released to enable reproducible benchmarking and foster culturally aligned moderation tools for Bangla online discourse.

Abstract

Online social media platforms are central to everyday communication and information seeking. While these platforms serve positive purposes, they also provide fertile ground for the spread of hate speech, offensive language, and bullying content targeting individuals, organizations, and communities. Such content undermines safety, participation, and equity online. Reliable detection systems are therefore needed, especially for low-resource languages where moderation tools are limited. In Bangla, prior work has contributed resources and models, but most are single-task (e.g., binary hate/offense) with limited coverage of multi-facet signals (type, severity, target). We address these gaps by introducing the first multi-task Bangla hate-speech dataset, BanglaMultiHate, one of the largest manually annotated corpus to date. Building on this resource, we conduct a comprehensive, controlled comparison spanning classical baselines, monolingual pretrained models, and LLMs under zero-shot prompting and LoRA fine-tuning. Our experiments assess LLM adaptability in a low-resource setting and reveal a consistent trend: although LoRA-tuned LLMs are competitive with BanglaBERT, culturally and linguistically grounded pretraining remains critical for robust performance. Together, our dataset and findings establish a stronger benchmark for developing culturally aligned moderation tools in low-resource contexts. For reproducibility, we will release the dataset and all related scripts.

Figures (4)

• Figure 1: An example of hateful comment with its English translation showing type, severity and target of hate.
• Figure 2: Distribution of the MultiHate dataset across different categories.
• Figure 3: Heatmap demonstrating the relationship between type of hate and severity.
• Figure 4: Heatmap demonstrating the relationship between type of hate and target.