SyntaxMind at BLP-2025 Task 1: Leveraging Attention Fusion of CNN and GRU for Hate Speech Detection

TL;DR

The paper tackles Bangla hate speech detection under BLP-2025 Task 1 (Subtasks 1A and 1B) by proposing a unified architecture that fuses BanglaBERT embeddings with parallel CNN and Bi-GRU branches, each equipped with attention, followed by a fusion layer and a final classifier. Empirical results show competitive performance, achieving $0.7345$ micro F1 in Subtask 1A (2nd place) and $0.7317$ micro F1 in Subtask 1B (5th place), demonstrating the effectiveness of attention-based feature fusion in capturing both contextual semantics and local linguistic cues. The approach addresses dataset-level challenges such as class imbalance and language-specific biases in pretrained models, highlighting the value of tailored preprocessing and Bangla-specific contextualization. The work suggests extending the framework to Subtask 1C to further evaluate its generalization across task variations and target groups.

Abstract

This paper describes our system used in the BLP-2025 Task 1: Hate Speech Detection. We participated in Subtask 1A and Subtask 1B, addressing hate speech classification in Bangla text. Our approach employs a unified architecture that integrates BanglaBERT embeddings with multiple parallel processing branches based on GRUs and CNNs, followed by attention and dense layers for final classification. The model is designed to capture both contextual semantics and local linguistic cues, enabling robust performance across subtasks. The proposed system demonstrated high competitiveness, obtaining 0.7345 micro F1-Score (2nd place) in Subtask 1A and 0.7317 micro F1-Score (5th place) in Subtask 1B.

Figures (2)

• Figure 1: Class distributions for BLP-2025 Task 1
• Figure 2: Proposed Model