FCProfiler: Structured and Deterministic Pipeline for Recipe-Level Nutrient Estimation

Abstract

Food and nutrition question answering involves resolving ambiguous ingredient terminology and diverse household measurement expressions, and converting them into representations compatible with nutrient databases. In FoodBench-QA, recipe-level nutrient estimation requires consistent handling of heterogeneous and imprecise measurement descriptions. We propose FoodComponentProfiler (FCProfiler), a deterministic pipeline that treats nutrient estimation as a structured measurement resolution problem. The pipeline is composed of multiple stages, including parsing, normalization, unit canonicalization, gram conversion, and nutrient estimation, with each step designed to remain transparent and traceable. Unit canonicalization combines rule-based standards with data-driven unit expansion from large-scale recipe corpora, enabling broader coverage of real-world measurement variations. Gram conversion grounds quantities in ingredient-specific portion information, enabling accurate and traceable mass computation. Experimental results show that accurate nutrient estimation mainly depends on reliable unit normalization and ingredient-specific measurement conversion. Additionally, FCProfiler achieves performance comparable to FoodyLLM, demonstrating that explicit measurement grounding serves as an effective alternative to implicit reasoning. The proposed methodology preserves interpretability while maintaining strong performance in food and nutrition question answering.

Resources

Details