Published in the IEEE ICHMS 2025 proceedings, pp. 168–173.
This paper studies how formal instruction structure affects an LLM's ability to perform complex data annotation. Instead of focusing solely on model selection, it addresses task engineering: decomposing requirements, clearly defining classes, context, and constraints, and improving reproducibility and quality.
This research builds on earlier dialogue-annotation studies and treats instruction engineering as a controllable mechanism for LLM-assisted annotation.