Embedding-Based Data Matching for Disparate Data Sources

Dealing with heterogeneous sources is an important challenge in the field of knowledge discovery and management. Schema matching methods are employed to solve this problem using three approaches: schema-based, instance-based, or a combination. This paper focuses on mapping between a schema-available (only) data source and a data source containing both schema and instance (both). Given the lack of suitable methods for aligning these two types of sources, we propose an approach using embedding models to provide vector modelling of sources and calculate similarities between data. Our solution consists in combining domain-specific embedding models and cross-domain embedding models to make data matching possible and efficient between the above-mentioned data sources. We have conducted several experiments using the Valentine datasets to evaluate our data matching method on several disparate tabular data. The result indicate effectiveness in terms of stability and ablation handling.