Feasibility of emotions as features for suicide ideation detection in social media

Suicide-related social media message detection is an important issue. Such messages can reveal a warning sign of suicidal behaviour. This paper examines the efficacy of using emotions as sole features to detect suicide-related messages. We investigated two methods which use a single emotion and a set of seven emotions as features respectively. For emotion classification, we used a classifier based on BERT named "Emotion English DistilRoBERTa-base". For detecting suicide-related messages, we tested Naive Bayes and Support Vector Machine. As our training/test data for suicide message detection, we used a publicly available dataset collected from Reddit in which each post is labelled as "suicide" or "non-suicide". Our method obtained accuracies of 76.2% and 76.8% for detecting suicide-related messages with Naive Bayes and Support Vector Machine respectively. Our experiment also shows that three emotion categories, "anger", "fear" and "sadness", have a strongest correlation with suicide-related messages.