LARGE LANGUAGE MODELS FOR AUTOMATED DATA MAPPING IN ENTERPRISE INFORMATION SYSTEMS

Data mapping mapping the relationship between disparate schemas of disparate sources and targets is one of the most time-consuming integration challenges in enterprise information system (EIS) integration, M&A and cloud migration efforts. Existing schema-matching approaches based on rules or machine learning are highly handcrafted, require labeled training pairs, and rely on relatively strict assumptions regarding the types of schemas in the respective domains, making it hard to generalize to different enterprise resource planning (ERP), customer relationship management (CRM), and legacy database systems. In this paper, we explore how to leverage large language models (LLMs) for enterprise data mapping automation and enhancement. To this end, we propose an end-to-end architecture which includes schema serialization, prompt based and fine-tuned LLM reasoning, confidence-based ranking, and human-in-the-loop validation to produce accurate mapping at the attribute level with minimal supervision. The proposed system was tested using a composite benchmark of 101 source schemas and 6,544 labelled mapping pairs from three data sets sourced from ERP, CRM and healthcare interoperability. The experimental results demonstrate that fine-tuned LLM with a mapping F1-score of 91.0% performs better than rule-based matching (58.7%), Random Forest classifiers (70.1%), BERT-based matching (76.9%) and zero-shot GPT-3.5 (79.8%). Moreover, we demonstrate that the accuracy of the maps obtained can be significantly increased using just a few few-shot examples, which demonstrates good sample efficiency. The results indicate that data mapping with LLMs can significantly lower manual data integration effort without compromising accuracy, paving the way for self-service, semantically-aware enterprise data integration.