SMAT: An attention-based deep learning solution to the automation of schema matching

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Management

Background:

• Schema matching is crucial for data integration but challenging.
• Existing methods are often low-quality or require extensive manual effort.
• Data privacy concerns, especially in healthcare, necessitate schema-level matching.

Purpose of the Study:

• To propose SMAT, a novel deep learning model for semantic schema matching.
• To address limitations of current schema matching techniques, including privacy concerns.
• To introduce a new healthcare-domain benchmark dataset for schema matching evaluation.

Main Methods:

• Developed SMAT, a deep learning model leveraging advanced NLP techniques.
• Utilized only attribute names and descriptions for semantic mapping.
• Created the OMAP benchmark dataset from real-world healthcare schema mappings.

Main Results:

• SMAT demonstrates potential for automating schema-level matching tasks.
• The model effectively learns semantic mappings without domain-specific encoding.
• Evaluations on benchmark datasets confirm SMAT's efficacy.

Conclusions:

• SMAT offers a privacy-preserving and easily deployable solution for schema matching.
• The model shows promise in advancing automated data integration.
• The OMAP dataset facilitates further research in healthcare data interoperability.