An Evaluation of Large Language Models for Supplementing a Food Extrusion Dataset.

Area of Science:

• Food Science and Technology
• Data Science
• Artificial Intelligence

Background:

• Food extrusion is a key industrial process for creating structured food products, but lacks standardized data.
• Existing research data is fragmented, hindering scientific synthesis, product development, and process optimization.
• Manually curated datasets offer high quality but are limited by time and resources.

Purpose of the Study:

• To introduce a manually curated food extrusion literature dataset.
• To propose and evaluate a method for supplementing this dataset using large language models (LLMs).
• To assess the accuracy and limitations of LLMs in extracting structured food extrusion data.

Main Methods:

• Manual curation of a food extrusion literature dataset, including publication details, product types, process parameters, formulation data, experimental variables, and characterization metrics.
• Development and application of LLM-based methods to extract structured data from scientific literature on food extrusion.
• Comparative analysis of LLM-extracted data against manually curated data to evaluate accuracy and identify challenges.

Main Results:

• A comprehensive, manually curated food extrusion dataset has been established.
• LLMs demonstrate significant capability in extracting structured data from food extrusion literature.
• Identified challenges with LLMs include hallucination and missing contextual details, necessitating human validation.

Conclusions:

• LLMs show promise as a tool for supplementing food extrusion datasets, offering substantial time savings.
• Human validation remains crucial to ensure the quality and accuracy of LLM-generated data.
• Leveraging LLMs alongside human curation presents a viable strategy for improving food extrusion data accessibility and utility.