NL-to-SQL Training Data Augmentation for Aggregation and Sorting

The patent applies preliminary action by performing data augmentation before the main training process. Synthetic training examples are generated in advance using template-based and keyword-based strategies, creating an expanded training dataset that incorporates various database operations like aggregation and sorting. This preliminary data preparation reduces the need for manually collected labeled data while maintaining model performance.

The patent uses copying by creating synthetic copies of existing training examples through template substitution and keyword replacement. Existing natural language-SQL pairs are replicated and modified by inserting different keywords and templates, generating new training examples that preserve the original structure while varying the content. This copying approach efficiently expands the training dataset without requiring new manual annotations.

The patent applies local quality by focusing training data augmentation on specific database operations that require higher accuracy, such as aggregation and sorting. Instead of uniformly treating all query types, the system identifies operations where precision is critical and applies targeted keyword-based and template-based augmentation strategies specifically for these operations. This localized approach improves accuracy for critical operations without unnecessarily increasing computational burden across the entire model.

The patent uses parameter changes by modifying training data characteristics rather than changing the model architecture or training algorithm. The data augmentation process alters parameters such as keyword variations, template structures, and example diversity in the training set. These parameter changes in the input data enable the model to learn better representations for specific operations like aggregation and sorting, improving accuracy without requiring changes to the computational model itself.