LGEQRE: Learning Guided Enumerative Synthesis for Query Reverse Engineering

Abstract To address the problem of users' lack of SQL query writing skills, Query Reverse Engineering (QRE) was proposed, where the goal of QRE is to generate a SQL statement based on a given database and query output table. SQUARES is one of the state-of-the-art models in the field, which enumerates constraint-compliant programs using a solver-based enumerator, and since the Solver randomly enumerates candidate programs, SQUARES synthesis is not very efficient. In this paper, we propose LGEQRE based on SQUARES, a learning-based approach to guide the enumeration of candidate programs. LGEQRE predicts the operators be required by neural network, sorts and deletes operators based on the prediction, and uses an Optimizer-based enumerator to enumerate programs according to the predicted probability of the operators. Under the same experimental conditions, the experimental results showed that LGEQRE increased the synthesis rate from 80% to 89.1% and reduced the average synthesis time from 251s to 117s compared to SQUARES.