Empirical Analysis for Detecting Arabic Online Suicidal Ideation

Suicidal ideation is a serious social health problem influenced by personal, social, and negative life events. Suicide ideation expressed in social media can provide insights into suicidal attempts. Detecting and tracking people at risk of suicide is a critical task. A massive amount of Arabic content has become available on online social media. Early diagnosis of mental illness, including suicidal thoughts, is challenging, particularly within the Arabic culture, due to stigma and a lack of awareness. Few recent studies have investigated the automatic detection of Arabic suicidal thoughts either because of the lack of a dataset or difficulties in dealing with the Arabic language. Therefore, our study investigates the performance of using several deep classifiers with Arabic pre-trained language models to detect suicide ideation using deep learning on the largest available dataset. We selected Universal-Sentence-Encoder-multilingual (USE) and MARBERT models for conducting the experiment work. The experiment results show that the USE and deep dense model achieved 82.6% F1score. For MARBERT, the CNN+LSTM outperforms other presented deep models, which achieved an 80.8% F1score.