ARE SQL for improving semantic alignment in Chinese multi round text to SQL parsing

Abstract Accurately maintaining semantic alignment across dialogue turns is crucial for multi-round Text-to-SQL parsing. This paper proposes ARE-SQL to address this challenge. The model features a state tracker to maintain dialogue context and a correlation enhancer to deepen the interaction between text and database schema. The state tracker refines the alignment matrix by inheriting and updating historical alignment information via an attention mechanism, while the correlation enhancer captures indirect semantic links through relation-aware self-attention. Experimental results on the CHASE dataset demonstrate that ARE-SQL outperforms RATSQL by 5.2% and EditSQL by 3.8% in terms of Question Match accuracy, and improves Interaction Match by 4.5% over IGSQL.