Integrating Relational Structures into Text-to-SQL

NetMind in cooperation with Shanghai Jiao Tong University has created an innovative “RASAT” or relation-aware self-attention transformer model that integrates relational architecture into a pre-trained sequence-to-sequence model.

We've applied this model to handle text-to-SQL tasks, which automatically translate a user's natural language questions into executable SQL queries. This automation has the potential to significantly break down barriers for non-expert users who wish to interact with databases. Industries with huge relational databases, like healthcare, financial services and sales, can make frequent use of such a tool.

The RASAT model inherits T5 Transformer, but the original self-attention modules in the encoder are substituted with relation-aware self-attention. Experiments have shown that RASAT can achieve state-of-the-art performances in three competitive benchmarks Spider, SParC, and CoSQL.