AyahVerse at MAHED Shared Task: Fine-Tuning ArabicBERT with Preprocessing for Hope and Hate Detection
Year: 2025
Type: conference-paper
Abstract: We participated in Subtask 1 of the MAHED Shared Task 2025, which focuses on detecting hope, hate, and not applicable labels in Arabic content.In this work, we tested a multiclass classifier for hope, hate, and not applicable label detection from a dataset provided by organizers as Subtask 1.We approach the task by two methods.The first one is a fine-tuned Arabic model, ArabicBERT, on a multiclass classification task.The second one is a two-step stacked architecture.Both of them include a dedicated pipeline for specific Arabic preprocessing with different techniques.Official results are 0.38 F1 score on the validation set and 0.47 on the test set with a single multiclass classifier.Post-submission improvements resulted in macro-F1 scores of 0.60(validation) and 0.63(test) for the single classifier, and 0.59(validation) and 0.91(test) for the stacked classifier.
Authors: Ibad-ur-Rehman Rashid, Muhammad Hashir Khalil
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W4416037605