Suicide Risk Assessment on Social Media with Semi-Supervised Learning

Area of Science:

• Computational linguistics
• Mental health informatics
• Machine learning for social good

Background:

• Social media platforms are increasingly used by individuals expressing suicidal ideation.
• Automated suicide risk assessment using natural language processing (NLP) is promising but hindered by insufficient labeled data and class imbalance.
• Existing methods struggle with the inherent challenges of real-world social media data for suicide risk detection.

Purpose of the Study:

• To develop a robust semi-supervised learning framework for automated suicide risk assessment.
• To address the challenges of limited labeled data and class imbalance in suicide-related social media content.
• To improve the accuracy and reliability of suicide risk detection models.

Main Methods:

• Proposed a semi-supervised framework combining 500 labeled and 1,500 unlabeled social media posts.
• Expanded the self-training algorithm with a novel pseudo-label acquisition process tailored for imbalanced datasets.
• Implemented manual verification of a subset of pseudo-labeled data to ensure quality and reliability.
• Evaluated multiple models, identifying RoBERTa as the optimal backbone for the framework.

Main Results:

• The semi-supervised framework significantly improved the model's ability to assess suicide risk.
• The novel pseudo-labeling strategy effectively handled class imbalance issues.
• Partially validated pseudo-labeled data, combined with ground-truth labels, enhanced predictive performance.
• RoBERTa demonstrated superior performance as the underlying model for risk assessment.

Conclusions:

• Semi-supervised learning, particularly with a refined pseudo-labeling process, is effective for suicide risk assessment on social media.
• The proposed framework offers a viable solution to data scarcity and imbalance problems in this critical domain.
• This approach holds potential for developing more accurate and scalable automated systems for suicide prevention.