Search papers, labs, and topics across Lattice.
This paper benchmarks a spectrum of models, from TF-IDF to LLMs, for estimating nutrient content from recipe text, a challenging task due to ambiguous language. They find that while TF-IDF offers speed, LLMs like Gemini 2.5 Flash, especially in a hybrid approach, achieve the highest accuracy by leveraging pre-trained knowledge to handle ambiguity and non-standard units. The study highlights a crucial trade-off between computational cost and accuracy for nutrient estimation in dietary monitoring.
LLMs crush traditional methods in recipe nutrient estimation, but the accuracy boost comes at a steep computational price, forcing a trade-off between precision and real-time performance.
Accurate nutrient estimation from unstructured recipe text is an important yet challenging problem in dietary monitoring, due to ambiguous ingredient terminology and highly variable quantity expressions. We systematically evaluate models spanning a wide range of representational capacity, from lexical matching methods (TF-IDF with Ridge Regression), to deep semantic encoders (DeBERTa-v3), to generative reasoning with large language models (LLMs). Under the strict tolerance criteria defined by EU Regulation 1169/2011, our empirical results reveal a clear trade-off between predictive accuracy and computational efficiency. The TF-IDF baseline achieves moderate nutrient estimation performance with near-instantaneous inference, whereas the DeBERTa-v3 encoder performs poorly under task-specific data scarcity. In contrast, few-shot LLM inference (e.g., Gemini 2.5 Flash) and a hybrid LLM refinement pipeline (TF-IDF combined with Gemini 2.5 Flash) deliver the highest validation accuracy across all nutrient categories. These improvements likely arise from the ability of LLMs to leverage pre-trained world knowledge to resolve ambiguous terminology and normalize non-standard units, which remain difficult for purely lexical approaches. However, these gains come at the cost of substantially higher inference latency, highlighting a practical deployment trade-off between real-time efficiency and nutritional precision in dietary monitoring systems.
