Tweet Classification to Assist Human Moderation for Suicide Prevention

Area of Science:

• Computational linguistics
• Artificial intelligence
• Mental health technology

Background:

• Social media platforms use interventions for suicide prevention, often relying on moderator review of self-harm content.
• Automated models struggle with nuanced language, like sarcasm, leading to potential misidentification of suicidal intent.

Purpose of the Study:

• To examine temporal language differences before suicidal intent expressions on Twitter.
• To develop and analyze time-aware neural models incorporating historical user language and emotional spectrum.
• To improve the identification of social media content indicative of suicidal intent.

Main Methods:

• Focused on Twitter posts with phrases similar to suicidal intent but potentially ambiguous.
• Analyzed temporal variations in general and emotional language preceding clear suicidal expressions.
• Developed and evaluated time-aware neural network models using historical user tweeting activity.

Main Results:

• The strongest model achieved high performance (macro F1=0.804, recall=0.813) in identifying social media content indicative of suicidal intent.
• Qualitative analysis of use cases demonstrated model decision-making processes.
• The study explored alleviating moderator burden within moderation constraints.

Conclusions:

• Data-driven models can significantly improve the detection of suicidal intent on social media.
• Ethical implications of using such models for inferring suicidal intent require careful consideration.
• These advancements offer potential support for suicide prevention initiatives on online platforms.