Extração de esquemas de dados semi-estruturados (JSON) utilizando Large Language Model (LLM)

Abstract

This study investigates the application of Large Language Models (LLMs), specifically Gemma 3-4B and Qwen 2.5-14B, for the automatic extraction of data schemas from JSON files. Considering the challenges imposed by the heterogeneity and dynamic nature of semi-structured data, an approach based on a pipeline of fragmentation, inference, and probabilistic schema fusion is proposed. The validation, performed using the Ajv validator on real-world and synthetic datasets, demonstrated success rates exceeding 99%, confirming the generalization capability of the approach. However, the experiments highlighted that the high computational cost and processing time of larger models still pose significant challenges to the scalability of the solution in resource-constrained environments.