AL-SQUARES: SQL Synthesis System with the Addition of Reducer

During the early stages of artificial intelligence development, researchers began exploring methods for computer-aided program synthesis. Among these methods, SQL reverse synthesis emerged as a crucial branch in the field of data analysis and became a research hot spot for applying artificial intelligence in the database domain. Program synthesis can be categorized into different types based on various requirements. This paper primarily focuses on input-output-based program synthesis, which involves generating programs based on given examples. SQUARES, an SQL reverse synthesizer proposed by scholars from the University of Lisbon, serves as the foundation of our work. We extend SQUARES by incorporating a reducer module based on attention-based RNN to guide program synthesis. With the integration of this module, the system is capable of removing less-relevant production rules, reducing the generation of irrelevant program statements, and effectively shrinking the program search space. Our model achieves an accuracy of 91.6% on the test set, while the baseline model only achieves 80.9%. This represents a significant improvement in success rate, with an increase of 10.7%.