An Integrated Interactive Framework for Natural Language to SQL Translation.

Numerous web applications rely on databases, yet the traditional database interface often proves inconvenient for effective data utilization. It is crucial to address the significant demand from a vast number of end users who seek the ability to input their requirements and obtain query results effortlessly. Natural Language (NL) Interfaces to Databases (NLIDBs) with interactive query mechanisms make databases accessible to end users and simultaneously retain user confidence in the results. This paper proposes an approach called IKnow-SQL for building interactive NLIDBs. IKnow-SQL introduces a unified framework for translation models to improve accuracy and increase interactivity. Specifically, IKnow-SQL first employs an underlying translation model to parse the semantics of a given NL query. By evaluating the model behavior, IKnow-SQL then recognizes the parts of the model output that may require human intervention. Next, IKnow-SQL presents clarifying questions to solicit and memorize user feedback until a polished result is obtained. Extensive experiments are performed to study IKnow-SQL on the public benchmark. The results show that the translation models can be effectively improved using IKnow-SQL with less user feedback.