Database Entity Recognition with Data Augmentation and Deep Learning (arxiv.org)

A recent study addresses the challenge of Database Entity Recognition (DB-ER) within natural language queries by introducing a comprehensive approach combining data augmentation and deep learning. The authors created a human-annotated benchmark derived from popular text-to-SQL datasets to evaluate DB-ER tasks. They also developed a novel data augmentation technique that automatically labels natural language queries based on corresponding SQL queries, significantly enriching the training data. The core of their approach is a specialized entity recognition model built on the T5 language model, fine-tuned for two distinct downstream tasks: sequence tagging and token classification. Experimental results show that this custom DB-ER tagger outperforms two leading state-of-the-art Named Entity Recognition (NER) systems in both precision and recall. Ablation studies reveal that the data augmentation strategy improves these metrics by over 10%, while fine-tuning the T5 model adds an additional 5-10% gain, underscoring the combined power of enriched training data and tailored model optimization. This work is significant for the AI/ML community as it provides a scalable, effective method for improving semantic parsing in text-to-SQL applications—an area crucial for making databases more accessible through natural language. By enhancing entity recognition accuracy, the method enables more precise query understanding, potentially advancing downstream tasks like automated SQL generation and conversational database interfaces.