Team R00 at Mowjaz Multi-Topic Labelling Task for Arabic Articles

Abstract

This paper describes the winning system for the Mowjaz Multi-Topic Labelling Task. The goal of the task is to classify articles based on their topics and predict multiple topics in one article. The proposed system is an ensemble model that consists of six BERT-Based models trained on different versions of the dataset. It achieved an F1-Micro score of 0.886 and an Accuracy score of 0.843 on the validation data. It also achieved an F1-Micro score of 0.8595 on the Test data, which led to ranking the model 1st in the Mowjaz Multi-Topic Labelling leaderboard. The current research work discusses the pre-trained language models used for the experimentation that led to the proposed system and shows the models’ performances on the Arabic Articles dataset.