Modeling SQL Statement Correctness with Attention-Based Convolutional Neural Networks
Pablo Rivas, Donald R. Schwartz · 2021 International Conference on Computational Science and Computational Intelligence (CSCI) · 2021
Automated grading of SQL statements is a topic of interest for instructors and students alike. It can give teachers additional time to be more effective in identifying issues quickly, and it can give students a preview of the grade they may receive. Existing attempts in the literature create models based on a variety of methodologies that exploit the structure of SQL statements and model answers; however, very few have leveraged the recent advances in deep learning. This paper employs a convolutional self-attention mechanism to learn complex contextual and grammatical dependencies directly from data of labeled SQL statements. Our experiments suggest that the proposed parameter-sharing strategy can adequately model the problem of detecting the correctness of an SQL statement with a balanced accuracy of up to 81.2% and an AUC of 0.87 in cross-validation.