Applications of NLP in recipe texts

Abstract

This study addresses the challenge of large-scale, multi-label recipe classification us- ing a real-world dataset of over 600,000 recipes collected from heterogeneous sources. The raw data exhibited significant noise, duplication, and label imbalance, motivating a comprehensive, multi-stage cleaning and preprocessing framework. Key steps included in- gredient normalization, instructions standardization, multi-label parsing, deduplication, and semantic category mapping into hierarchical supercategories. For modeling, we im- plemented a modular pipeline combining TF-IDF feature extraction, classical classifiers, XGBoost, and fine-tuned BERT models to capture both statistical and contextual signals. By adopting a per-supercategory strategy, we minimized cross-domain interference and achieved strong performance, with the fine-tuned BERT classifier attaining a weighted F1-score of 0.7996 and high accuracy on dominant labels. This work demonstrates how rigorous data preparation and modular modeling can enable fine-grained, interpretable recipe classification at scale, providing a robust foundation for downstream culinary ap- plications such as personalized meal planning and intelligent search.