Transformer Models for Authorship Profiling in Arabic Social Media Texts

Abstract

Authorship analysis is generally performed to extract information about the author or authors of documents by examining the inherent features present within these texts. The significance of the authorship profiling task is escalating, particularly with the widespread rise in social media users and platforms. This research explores Multi-Task Learning (MTL) with transformers, focusing on refining and evaluating advanced pre-trained models for author profiling using Arabic tweets. ARBERT, MARBERT, AraBERT, and BERT base Arabic transformers undergo fine-tuning for binary classification tasks in Arabic tweet analysis. The study incorporates MTL by concurrently training models on tasks such as dialect identification, sentiment analysis, and topic classification to enhance overall performance. Parameter optimization plays a pivotal role, achieving reliable results with AraBERT demonstrating the highest F1 score on the test dataset. MTL integration showcases promising outcomes, reinforcing the transformers' efficacy in authorship profiling.