Translating natural language questions to SQL queries (nested queries)

Real world questions are generally complex and need the user to extract information from multiple tables in a database using complex SQL queries like nested queries. Though the overall accuracy in translation of Natural Language queries to SQL queries lies close to 75%, the accuracy of complex queries is still quite less, around 60% in the current state-of-the-art models. In this vein, this study proposes to improve the current IRNet framework for translating natural language queries to nested SQL queries, one type of complex queries. Data oversampling is first used to boost the representation of nested queries in order to achieve this goal. Second, a novel loss function that computes the overall loss while accounting for the complexity of SQL, as measured by the quantity of SELECT columns and keywords in the SQL query. The proposed method exhibited a 5% improvement in prediction of hard and extra hard queries when tested on Spider's development dataset.