A Study on Machine Learning Techniques for the Schema Matching Networks Problem

Abstract

Schema Matching is the problem of finding semantic correspondences between elements from different schemas. This is a challenging problem, since the same concept is often represented by disparate elements in the schemas. The traditional instances of this problem involved a pair of schemas to be matched. However, recently there has been a increasing interest in matching several related schemas at once, a problem known as Schema Matching Networks, where the goal is to identify elements from several schemas that correspond to a single concept. We propose a family of methods for schema matching networks based on machine learning, which proved to be a competitive alternative for the traditional matching problem in several domains. To overcome the issue of requiring a large amount of training data, we also propose a bootstrapping procedure to automatically generate training data. In addition, we leverage constraints that arise in network scenarios to improve the quality of this data. We also propose a strategy for receiving user feedback to assert some of the matchings generated, and, relying on this feedback, improving the quality of the final result. Our experiments show that our methods can outperform baselines reaching F1-score up to 0.83.