Knowledge-Infused Dynamic Embedding for Predicting the Severity of Suicidal Ideation in Social Media

Depression accounts for an increased number of suicides in the United States. To identify those at risk and notify the authorities, social media can play an invaluable role. Posts made on social media can be collected to train predictive models capable of identifying individuals at risk. However, studies with an unsupervised method of data labeling are susceptible to misidentify individuals presenting signs of depression from those who provide support. Moreover, these studies tend to classify depression as a binary outcome, not considering its severity levels. This necessitates the use of accurate and efficient modeling techniques for the existing scarce but professionally annotated datasets. While pre-trained embeddings proved to be an efficient method of text representation, not all may fully encode emotions and sentiment polarities present in mental health related posts. In this work, we propose the use of a dynamic embedding infused with emotion and polarity knowledge for a more accurate representation of depression severity and suicidal ideation. The emotion, polarity, and context-aware generated embeddings are then utilized by bidirectional RNN (GRU and LSTM) model to provide the most accurate predictions. The results show that knowledge-infused dynamic embeddings with bidirectional LSTMs leads to an 8% increase in the overall AUC with improved prediction accuracy of suicidal ideations at its highest level.