Reproduce “In-context learning capabilities of Large Language Models to detect suicide risk among adolescents from speech transcripts” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Filomene Roquefort, Alexandre Ducorroy, Rachid Riad
Researchers from Callyope developed an automated system leveraging Large Language Models (LLMs) and in-context learning to detect suicide risk among Chinese adolescents from speech transcripts. This method achieved 0.68 accuracy on the test set, placing 3rd in the SpeechWellness Challenge, by focusing on linguistic cues when speech anonymization hindered acoustic analysis.
Key insights
- In-context learning significantly improves LLM classification accuracy for suicide risk detection, with the number of examples in the prompt being a statistically significant factor.
- Linguistic features derived from speech transcripts can effectively identify suicide risk even when acoustic markers are compromised by anonymization, outperforming acoustic-only baselines.
- Programmatic prompt engineering via DSPy provides a robust and reproducible approach for developing LLM-based systems for sensitive mental health applications.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.