Applied Intelligence Artificial Intelligence

Systematic optimization of a hybrid text–nutrient retrieval pipeline for food product matching

Svetlana Schmidt, Linda Klasen, Ute Nöthlings, Rafet Sifa
Springer Nature · 2026-11-01 · DOI 10.1007/s10489-026-07579-2

Abstract

Abstract Accurate dietary intake assessment is essential for research in nutritional epidemiology, public health, and sustainable nutrition. Integrating new processed foods into databases often remains a time-consuming manual task for dietitians. To support expert workflows, we propose a constraint-based information retrieval system that assists recipe matching by identifying nutritionally compatible database items with similar ingredient composition and product characteristics. The proposed approach combines lexical similarity, nutrient profile comparison, and structured product attributes within a two-stage retrieval architecture. Domain knowledge is incorporated through nutrient tolerance constraints, ingredient consistency, brand and category compatibility, enabling knowledge-driven candidate filtering prior to ranking. Building on a category-filtered baseline, we systematically refine parameters and introduce threshold-based candidate selection and confidence-aware brand filtering to refine candidate selection. Evaluation on a dietitian-annotated dataset demonstrates substantial improvements in precision among retrieved items while maintaining practical coverage, thereby enhancing retrieval reliability and efficiency for expert-assisted expansion of the food database.

Read at publisher →