High Value Payment System Data Inquiry Using a NL2SQL Framework

When applying the popular deep learning based NL2SQL models directly in a specific scenario, problems arise due to the characteristics rooted in background knowledge. In our case, the terminologies and abbreviations in the high value payment system database are the main obstacles. In this paper, a framework that is compatible with BERT-CN and RAT-SQL is proposed for data inquiry tasks within the high value payment system, in which both BERT and RAT-SQL are state of the art models achieved great performance in many tasks. Besides that, NER and data preprocessing toolkits are introduced to align the terminologies and abbreviations with the columns and tables. Both the training and testing stages show acceptable results and the reasons are well discussed. This framework has great potential to be extended to other application scenarios with minimal modifications.