Sefamerve R&D at ICICS 2021 Mowjaz Multi-Topic Labelling Task

Abstract

This paper describes our contribution to ICICS 2021 Mowjaz Multi-Topic Labelling Task. The purpose of the task is to classify Arabic articles based on their topics. Participating systems are expected to select one or more of the determined topics for each given article. In our system, we experiment with state-of-art pre-trained language models (GigaBERT-v4 and Arabic BERT) and a classical logistic regression to find the best effective model for the problem. We obtained the highest F1-score of 0.8563 with GigaBERT-v4 while Arabic-BERT and logistic regression reached 0.8442 and 0.8081 respectively. Our system ranked 2nd in the competition very close to the winner.