Detecting critical responses from deliberate self-harm videos on YouTube

YouTube is one of the leading social media platforms and online spaces for people who self-harm to search and view deliberate self-harm videos, share their experience and seek help via comments. These comments may contain information that signals a commentator could be at risk of potential harm. Due to a large amount of responses generated from these videos, it is very challenging for social media teams to respond to a vulnerable commentator who is at risk. We considered this issue as a multi-class problem and triaged viewers' comments into one of four severity levels. Using current state-of-the-art classifiers, we propose a model enriched with psycho-linguistic and sentiment features that can detect critical comments in need of urgent support. On average, our model achieved up to 60% precision, recall, and f1-score which indicates the effectiveness of the model.