Abstract
The screening of suicidal ideation on social media is a serious element of the proactive intervention of a mental disorder, especially in linguistically heterogeneous areas like Kazakhstan. Conventional text-only models are usually incapable of dealing with the vagueness of colloquialism and the absence of behavioural context. In the current paper, there is a proposal of a multimodal neural network architecture that combines multilingual BERT-base-cased Kazakh and Russian text embeddings with auxiliary temporal-engagement and demographic streams of data. The framework applies Bi-directional Long ShortMemories (BiLSTM) network, which is improved with an attention mechanism focusing on prioritizing high-risk linguistic features and abnormal social behaviors. It was tested and trained in a curated dataset of 2,500 posts of 1,200 respondents in five major regions of Kazakhstan. The results of the experiment indicate that the multimodal approach performs much better than the unimodal baselines, with the accuracy of 0.95, F1-score of 0.94 and Area Under the Curve (AUC) of 0.97. These results indicate that incorporation of social engagement measures and time features offer crucial context to resolve linguistic ambiguities and is a very powerful tool to detect early cases of suicidal ideation in Central Asian social media scenarios.