Enhancing Tabular Data Generation through Data and Knowledge Dual-Driven Approaches

Tabular data, as a fundamental medium for information storage in everyday contexts, is crucial for training various deep learning models. Despite its importance, current research predominantly addresses the processing and application of tabular data, with relatively little focus on the effective generation of high-quality tabular data, especially under constrained conditions. This study aims to address this gap by developing a high-quality tabular data generation method driven by both data and knowledge. To enhance the table generation capabilities of large models, we employ fine-tuning to improve table comprehension and integrate external knowledge bases to elevate the model's understanding from a knowledge perspective. Experimental results demonstrate that this dual-drive approach improves the accuracy of generating tabular flight data in the aviation sector by 22% compared to a baseline large language model.