A Transfer Learning Based Approach to Automatic SQL Translation

Abstract

Limitations of Oracle database in business support, development, operation, and maintenance hinder marketing 2.0's growth. Translating Oracle SQL statements to domestic databases is vital. A neural machine translation model based on migration learning resolves the lack of parallel corpus between Oracle and domestic databases. Encoder and decoder parameters of both databases' models are initialized using a trained Oracle database-side encoder and domestic database-side decoder. Fine-tuning and optimization produce high-quality parameters, improving translation performance. Results showed that the bilingual evaluation understudy (BLEU) of model reached 31.20 and an execution accuracy value of 84.3, outperforming the Transformer model by 13.28 and 13.3, respectively. Demonstrate support for database migration.