Gradient Masters at BLP-2025 Task 1: Advancing Low-Resource NLP for Bengali using Ensemble-Based Adversarial Training for Hate Speech Detection

2025-11-23

AI Summary

This work addresses hate speech detection in low-resource Bangla YouTube comments. We propose a multi-stage fine-tuning framework built upon a pre-trained Bangla language model, integrating heterogeneous models into a hybrid ensemble and incorporating adversarial training to enhance robustness and generalization. The method jointly tackles two fine-grained subtasks: hate speech classification (Subtask 1A) and target group identification (Subtask 1B). Evaluated on SemEval-2024 Task 12, our approach achieves 73.23% micro-F1 on Subtask 1A (6th place) and 73.28% micro-F1 on the more challenging Subtask 1B (3rd place), substantially outperforming baseline systems. Our key contribution is the first synergistic application of ensemble learning, adversarial training, and domain-adaptive fine-tuning to low-resource Bangla hate speech detection—effectively mitigating overfitting and distributional shift induced by data scarcity.

Technology Category

• Natural Language Processing: Fact-Checking / Misinformation Detection (NLP Focus)
• Machine Learning: Adversarial Learning & Robustness
• Multiagent Systems: Adversarial Agents

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Large pretrained models with web data
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

This paper introduces the approach of "Gradient Masters" for BLP-2025 Task 1: "Bangla Multitask Hate Speech Identification Shared Task". We present an ensemble-based fine-tuning strategy for addressing subtasks 1A (hate-type classification) and 1B (target group classification) in YouTube comments. We propose a hybrid approach on a Bangla Language Model, which outperformed the baseline models and secured the 6th position in subtask 1A with a micro F1 score of 73.23% and the third position in subtask 1B with 73.28%. We conducted extensive experiments that evaluated the robustness of the model throughout the development and evaluation phases, including comparisons with other Language Model variants, to measure generalization in low-resource Bangla hate speech scenarios and data set coverage. In addition, we provide a detailed analysis of our findings, exploring misclassification patterns in the detection of hate speech.

Problem

• Developing ensemble-based adversarial training for Bengali hate speech detection
• Addressing low-resource NLP challenges in Bangla language YouTube comments
• Improving classification of hate types and target groups in Bengali content

Innovation

• Ensemble-based fine-tuning strategy for hate speech
• Hybrid approach on Bangla Language Model
• Adversarial training for low-resource NLP scenarios