You Only Read Once (YORO): Learning to Internalize Database Knowledge for Text-to-SQL
This addresses efficiency and knowledge utilization issues for text-to-SQL applications, representing a novel paradigm shift rather than incremental improvement.
The paper tackles the problem of high inference cost and overlooked database knowledge in text-to-SQL systems by proposing YORO, a paradigm that internalizes database knowledge during training to eliminate schema encoding at inference, reducing input tokens by 66%-98% while achieving competitive performance on benchmarks and outperforming on large databases.
While significant progress has been made on the text-to-SQL task, recent solutions repeatedly encode the same database schema for every question, resulting in unnecessary high inference cost and often overlooking crucial database knowledge. To address these issues, we propose You Only Read Once (YORO), a novel paradigm that directly internalizes database knowledge into the parametric knowledge of a text-to-SQL model during training and eliminates the need for schema encoding during inference. YORO significantly reduces the input token length by 66%-98%. Despite its shorter inputs, our empirical results demonstrate YORO's competitive performances with traditional systems on three benchmarks as well as its significant outperformance on large databases. Furthermore, YORO excels in handling questions with challenging value retrievals such as abbreviation.