Pipeline and deep learning approach for NLIDB: A comparative study

Abstract

Databases are integral part of current world’s scenario of rich technology. Greater amount of the data in the world is stored in the databases. That amount of data storages can be utilized for various purposes in data science world. Besides potential usage and benefits of available data amounts, the requirement of formal language to access the databases is a huge hurdle. Structured Query language (SQL) is one of such formal languages to access the database. Besides its impact and powerful as a language it is not a common knowledge. Therefore, domain experts of some particular databases cannot access their data freely and easily. Web interfaces to access that data has their own limitation and do not fulfil the purpose to the maximum of the potential of data. Natural Language Interface to Database (NLIDB) system is natural solution for such problems. Text to SQL task in NLIDB system is being experimented with since 70s. Previously it was based on integrated methods and techniques from Natural Language Processing (NLP) and Data Science areas, those integrated frameworks generally known as pipeline methods. Recently, machine learning showed promising performance for the solutions to semantic problems. Which is why, deep learning had been adopted for text to SQL task as well. Currently NLIDB systems research is going on with both of the approaches of pipeline methods and deep learning methods in parallel. It is important at this time to analyze the latest work with both approaches and compare and identify their unique challenges and issues as well as findings and potential of both approaches for the NLIDB systems. In this paper, a comparative analysis is presented to find out the achievements and issues of NLIDB with pipeline methods and with deep learning methods regarding each of them.