Hybrid Embedding and Transformer-Based Architecture for Data Schema Transformation

In modern data-driven environments, heterogeneous data schemas present major challenges for ETL pipelines and seamless data exchange. To address this, a novel hybrid architecture is introduced: combining embedding-based methods (SBERT, Word2Vec) with a transformer detector, linked through an XOR-based embedding fusion mechanism. Built on a custom data set of more than 10,000 data fields in domains such as ecommerce, healthcare, and payment systems. The approach captures semantic and contextual nuances through aliases, metadata, types, and descriptive attributes. The fusion layer filters redundant signals while preserving key semantic features, enabling more accurate schema mapping. By refining fused embeddings with a Transformer Decoder, the architecture automates field-to-field alignment, reducing manual intervention and errors. This method demonstrates adaptability to unseen formats and potential for streamlining complex schema transformation tasks.