Predicting Suicide Ideation from Social Media Text Using CNN-BiLSTM

Predicting suicide ideation is crucial for mental health assessment, especially as clinical methods have been unsuccessful due to victims’ reluctance to seek help. Also, due to the time-consuming nature of clinicians having to review patient case notes individually, there is a risk that victims might commit suicide before being detected. Deep learning models such as the Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) models have shown promise in improving suicide risk assessment. A key challenge, however, is the need to find the right combination of word embeddings for vectorizing texts for these Deep-learning methods. This work uses a deep learning network composed of CNN and Bidirectional LSTM layers with two different word embedding techniques, Word2Vec and FastText. Using Word2Vec as the baseline word embedding. Experiments on a Reddit dataset of 232,074 posts gave test set F1-scores of 94% using FastText and 90% using Word2Vec. It was observed that FastText could give better performance with less overfitting than Word2Vec.