CairoDep: Detecting Depression in Arabic Posts Using BERT Transformers

Early detection and intervention of mental illness can significantly improve the chances of successful treatment and cure. With the wide adoption and use of social media platforms, the posts of someone give a window to his or her mind that can help detect mental health issues. Numerous researches were done to identify mental health conditions from posts in English language. Very few were done for Arabic language using traditional natural language processing (NLP) models. In this novel research, we use the latest NLP models, namely Bidirectional Encoded from Representations Transformers (BERT), for detecting depression in Arabic social media posts. We built CairoDep, a set of models and a benchmark dataset for this purpose. First, we built CairoDep v1.0 dataset, a labeled dataset of 7000 posts including 3400 normal (non-depressed) posts and 3600 depressed posts. Our dataset was collected from multiple sources, specifically crowdsourcing, Arabic mental health forums and pages, readily available datasets and translation from English datasets. Second, we further trained two pre-trained BERT transformers for Arabic language to detect depression from Arabic posts in Modern Standard Arabic (MSA) and dialectic Arabic, namely ARABERT and MARBERT. We further trained and evaluated them on the dataset using 80%-20% split. We achieved accuracy, precision, recall and F1-Score values of 96.93%, 96.92%, 96.93% and 96.92% for ARABERT and of 96.07%, 96.11%, 96.04% and 96.07% for MARBERT respectively. These results are quite superior to the results reported in the literature for depression detection using lexicon analysis or traditional machine learning techniques. They open the door for deploying NLP transformers to develop mental health AI-powered applications and platforms for Arabic speakers. We developed these models as part of iHayaNow, an AI-based mental health helpline and a holistic platform for Arabic speakers that is under research and development.