RECONSTRUCTING ENTITY RELATIONSHIPS IN DATABASE SCHEMAS WITH PLANTUML AND LLMS

Abstract

The article explores the potential of using Large Language Models (LLMs) for automatically restoring relationships between tables in SQL databases with incompletely defined foreign keys. To evaluate the ability of LLMs to infer foreign keys from textual descriptions of table structures, an experimental database was created. The database schema, excluding relationships, was provided as input to two large language models: ChatGPT-4o and Claude 3.7 Sonnet. For analysis purposes, only basic information was provided to the LLMs: table names, field names, and primary keys, without any data examples. The ChatGPT-4o model successfully detected all relationships between tables but demonstrated limitations in determining the types of these relationships: all were classified as “one-to-one”, regardless of their actual structure. This indicates the model's inability to accurately interpret the type of relationships based on textual descriptions. In contrast, the Claude 3.7 Sonnet model not only correctly identified all existing relationships, but also correctly determined their types (e.g., one-to-many), demonstrating higher accuracy and a deeper understanding of the database structure within the task at hand. The description of the table structure was provided to the language models in PlantUML format, ensuring a standardized, clear and unambiguous representation of the input data. Based on the modeling results, ER diagrams were also constructed in PlantUML format. The experiment confirms the effectiveness of LLMs in reconstructing missing foreign keys and shows potential for automated analysis, documentation, and improvement of existing databases. Following consistent naming conventions during schema design significantly simplifies both the work of developers and the automated processing of database structures by intelligent systems, playing a crucial role in these processes.