LLM-Matcher: A Name-Based Schema Matching Tool using Large Language Models

Abstract

We present LLM-Matcher, an interactive name-based schema matching system that utilizes large language models (LLMs) to identify correspondences between source and target schema elements relying solely on their names and descriptions. This tool is specifically designed for restricted environments where instance-based schema matching is not possible, such as the healthcare domain where instance access is often prohibited. LLM-Matcher is based on an extensive experimental study, showing the capabilities of LLMs in the schema matching task. Our system is specifically tailored towards users with sufficient domain knowledge and offers an interpretable initial mapping that can be further refined by providing textual feedback. This feedback allows to rectify model misconceptions as well as improve the quality of schema element descriptions. This paper provides a comprehensive overview of LLM-Matcher, explores its application in the healthcare domain, allows users to gain additional insight into our experimental study, and outlines different steps showcased in the demonstration.