Improving Suicidal Ideation Detection in Social Media Posts: Topic Modeling and Synthetic Data Augmentation Approach

Area of Science:

• Computational Social Science
• Artificial Intelligence
• Public Health Informatics

Background:

• Online suicide discussions are crucial for public health but often lack representation of marginalized communities.
• Underrepresentation in social media data leads to underperformance of AI models on critical topics.

Purpose of the Study:

• To analyze online suicide discussions and identify underrepresented topics.
• To generate synthetic data to improve the diversity and coverage of suicide-related datasets for AI training.

Main Methods:

• Unsupervised and guided topic modeling on social media data.
• Scoping review of psychology literature for suicide risk factors.
• Generative large language models (GPT-3.5 Turbo) for synthetic data creation.
• Evaluation of synthetic data for readability, complexity, and utility.

Main Results:

• Critical topics, especially those concerning marginalized communities and racism, were underrepresented in real social media data.
• Synthetic data generation improved topic diversity and augmented datasets.
• Fine-tuning AI classifiers with augmented data significantly improved suicidal ideation detection accuracy (F1-score from 0.70 to 0.90).

Conclusions:

• Synthetic datasets are valuable for understanding online suicide discourse.
• AI models trained with synthetic data augmentation show improved accuracy in detecting suicidal narratives.