[Paper Review] FEAST: Retrieval-Augmented Multi-Hierarchical Food Classification for the FoodEx2 System

Lorenzo Molfetta, Alessio Cocchieri+4 more|arXiv (Cornell University)|Mar 3, 2026

Nutritional Studies and Diet0 citations

TL;DR

FEAST introduces a retrieval-augmented, multi-stage pipeline that maps free-text food descriptions to FoodEx2 codes by sequentially identifying base terms, facet categories, and facet descriptors, using dense retrieval, reranking, and optionally LLMs, to address hierarchical and data-sparse label spaces.

ABSTRACT

Hierarchical text classification (HTC) and extreme multi-label classification (XML) tasks face compounded challenges from complex label interdependencies, data sparsity, and extreme output dimensions. These challenges are exemplified in the European Food Safety Authority's FoodEx2 system-a standardized food classification framework essential for food consumption monitoring and contaminant exposure assessment across Europe. FoodEx2 coding transforms natural language food descriptions into a set of codes from multiple standardized hierarchies, but faces implementation barriers due to its complex structure. Given a food description (e.g., "organic yogurt''), the system identifies its base term ("yogurt''), all the applicable facet categories (e.g., "production method''), and then, every relevant facet descriptors to each category (e.g., "organic production''). While existing models perform adequately on well-balanced and semantically dense hierarchies, no work has been applied on the practical constraints imposed by the FoodEx2 system. The limited literature addressing such real-world scenarios further compounds these challenges. We propose FEAST (Food Embedding And Semantic Taxonomy), a novel retrieval-augmented framework that decomposes FoodEx2 classification into a three-stage approach: (1) base term identification, (2) multi-label facet prediction, and (3) facet descriptor assignment. By leveraging the system's hierarchical structure to guide training and performing deep metric learning, FEASTlearns discriminative embeddings that mitigate data sparsity and improve generalization on rare and fine-grained labels. Evaluated on the multilingual FoodEx2 benchmark, FEAST outperforms the prior European's CNN baseline F1 scores by 12-38 % on rare classes.

Motivation & Objective

• Motivate the challenge of hierarchical text classification and XML in the real-world FoodEx2 taxonomy.
• Propose FEAST to decompose FoodEx2 coding into base term, facet category, and descriptor prediction.
• Leverage hierarchical structure and deep metric learning to mitigate data sparsity and improve rare-label generalization.
• Demonstrate retrieval-augmented methods with flexible reranking and LLM-based options across stages.

Proposed method

• Represent FoodEx2 as a three-stage pipeline: base term identification, facet category classification, and facet descriptor assignment.
• Use a retrieval-augmented framework with a bi-encoder for embedding and a cross-encoder reranker to refine top candidates.
• Apply hard negative mining derived from the taxonomy to train robust embeddings.
• Employ retrieval-and-rerank components with options for LLM-based classifiers to map NL descriptions to code components.
• Train an LLM (LLaMA-3.1-8B) with LoRA adapters to jointly handle all three sub-tasks in a multitask setup.
• Evaluate using retrieval metrics (Accuracy@K, Recall@K, NDCG, MRR, MAP) and classification metrics (Precision, Recall, F1) across tasks.

Experimental results

Research questions

• RQ1 How can retrieval-augmented methods improve base term, facet category, and facet descriptor predictions in FoodEx2?
• RQ2 Can a three-stage FEAST pipeline outperform CNN-based baselines on rare and fine-grained FoodEx2 codes?
• RQ3 What is the impact of hard negative mining and hierarchical structure on generalization to unseen base terms and facets?
• RQ4 What are the trade-offs between encoder-based reranking and LLM-based classification in this hierarchical, multi-task setting?

Key findings

• FEAST outperforms the prior European CNN baseline by 12–38% on rare classes in FoodEx2 coding.
• Retrieval components achieve strong base term and descriptor retrieval (e.g., Base Term Acc@1 96.57%, Descriptor Acc@1 98.90%).
• Rerankers improve base-term accuracy to 91.01% and descriptor accuracy to 96.03% @1.
• Bi-encoder classifiers show strong but uneven performance across facet categories, with DeBERTa-v3-base achieving Task II metrics around 73% @1 and higher at ranks 3–10.
• LLM-based classification (LLaMA-3.1-8B with LoRA) excels at Task I and Task III but shows mixed results for Task II, suggesting category-level multitask benefits require careful prompting and/or model choice.
• Across tasks, FEAST achieves near-perfect Task I and Task III accuracy in validation settings and demonstrates robustness under out-of-sample evaluation.