Abstract
Relational database management systems (RDBMSs) are powerful because they areable to optimize and answer queries against any relational database. A naturallanguage interface (NLI) for a database, on the other hand, is tailored tosupport that specific database. In this work, we introduce a general purposetransfer-learnable NLI with the goal of learning one model that can be used asNLI for any relational database. We adopt the data management principle ofseparating data and its schema, but with the additional support for theidiosyncrasy and complexity of natural languages. Specifically, we introduce anautomatic annotation mechanism that separates the schema and the data, wherethe schema also covers knowledge about natural language. Furthermore, wepropose a customized sequence model that translates annotated natural languagequeries to SQL statements. We show in experiments that our approach outperformsprevious NLI methods on the WikiSQL dataset and the model we learned can beapplied to another benchmark dataset OVERNIGHT without retraining.