Schema Matching Method Using Large Language Models and Graph Path Search (an Industrial Case Study of Production Planning System Integration)

One key challenge in the system integration business is improving the efficiency of schema matching. Schema matching is a task where two different schemas are given as input, and semantically corresponding columns are identified. Recent methods using large language models (LLM) outperform rule-based and deep-learning approaches. However, the LLM-based approaches face two challenges: improving matching accuracy of (i) one-to-many cardinality and (ii) foreign keys (FKs). For the first challenge, we focused on the features that the one-to-many or one-to-zero cardinalities can be excluded if the column is an identifier (ID) or essential for the target application (i.e. production planning system), respectively, and thus the possible multiplicity can be classified into four patterns. Based on the features, we proposed a decomposition approach where an original problem is divided into four subproblems according to the four multiplicity patterns, and we use prompts specialized for each subproblem. For the second challenge, we focused on the features that database schemas can be modeled as graphs, and the FKs can be identified in the shortest path between the primary keys (PKs); thus, FKs can be identified by graph path search between PKs. Based on the features, we proposed a graph-based approach where PKs are matched first using LLM, and each FK is matched later using graph path search between corresponding PKs. The effectiveness of our proposed approaches was demonstrated through an industrial case study.