Recent advances in large language models (LLM) offered human-like capabilities for comprehending emotion and mental states. Prior studies explored diverse prompt engineering techniques for improving classification performance, but there is a lack of analysis of prompt design space and the impact of each component. To bridge this gap, we conduct a qualitative thematic analysis of existing prompts for emotion and mental health classification tasks to define the key components for prompt design space. We then evaluate the impact of major prompt components, such as persona and task instruction, on classification performance by using four LLM models and five datasets. Modular prompt design offers new insights into examining performance variability as well as promoting transparency and reproducibility in LLM-based tasks within health and well-being intervention systems.
https://dl.acm.org/doi/10.1145/3706598.3713888
The ACM CHI Conference on Human Factors in Computing Systems (https://chi2025.acm.org/)