A Composite Schema Matching Method Based on Contrastive Learning in Data Integration

Schema matching is a crucial step in data integration for in-depth information mining, aimed at capturing seman-tic correspondences between elements of multi-source datasets. The existing schema matching methods still face problems such as insufficient representations of schema information diversity, poor cross-domain expression ability, and inadequate adaptability to complex data environments. We propose a composite schema matching method based on contrastive learning. Our method de-signs multi-domain word lists for data augmentation, the schema-based matching part fine-tunes pre-trained models through con-trastive learning and finally designs multi-model voting strategies to complete the composite matches. It also designs a format validation strategy to enhance the performance of the schema matching model. The experimental results show that our method has improved the F1 comprehensive performance indicators on three benchmark datasets by an average compared to existing benchmark methods.