Suicidal Thoughts Detection on Social Media in Roman Urdu Using an Attention‐Based Hybrid Deep Learning Model

ABSTRACT Early detection of suicidal ideation on social media is crucial for timely intervention and prevention. In Pakistan and other Asian countries, a significant number of users communicate on social media in Roman Urdu, a non‐standardized script of Urdu, which poses challenges for text analysis. Prior studies indicate that there is no significant contribution in predicting suicide via Roman Urdu because most research has primarily focused on formal and structural languages, leaving a gap in research for low‐resourced languages. The proposed study developed a novel dataset of Roman Urdu user‐generated suicidal comments collected from YouTube, Facebook, and Reddit. BERT embedding has been used as an input feature to acquire rich semantic representations. We propose a hybrid deep learning‐based approach that combines CNN, LSTM‐GRU layers, and an attention mechanism, which is evaluated against various baseline models. Experimental findings reveal that the proposed model achieved an accuracy of 96%, which outperforms all baselines in the recognition of suicidal content in Roman Urdu text.