NguyenTriet at MAHED Shared Task: Ensemble of Arabic BERT Models with Hierarchical Prediction and Soft Voting for Text-Based Hope and Hate Detection
Year: 2025
Type: conference-paper
Abstract: We present the NguyenTriet system for the MA-HED 2025 shared task on multimodal detection of hope and hate emotions in Arabic content (Zaghouani et al., 2025).The challenge was divided into three subtasks: text-based hate and hope speech classification in Arabic text; multitask emotion, offensive language, and hate detection in Arabic text with a hierarchical structure; and detecting hateful memes from multimodal text-image pairs.Our participation focused on Subtasks 1 and 2. For Subtask 1, we employed an ensemble of Arabic BERT models for multi-class classification.In Subtask 2, we implemented a hierarchical classification framework utilizing a similar ensemble methodology, where emotion predictions are leveraged through a cascaded pipeline architecture to inform downstream hate and offensive detection tasks.Our approach achieved macro-F 1 scores of 0.707 (3rd place) on Subtask 1 and 0.553 (2nd place) on Subtask 2.
Authors: Nguyen Minh Triet, Thìn Đặng Văn
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037762