EviLink: Multi-Path Schema Linking with Uncertainty-Guided Evidence Acquisition for Large-Scale Text-to-SQL

EviLink has been introduced as a novel approach to schema linking in large-scale Text-to-SQL systems, addressing the challenges of identifying relevant schema contexts from extensive and ambiguous databases. This method reframes schema linking as uncertainty-aware schema-need inference, allowing for the acquisition of evidence only where necessary.

The development of EviLink is significant as it enhances the balance between schema completeness, relevance, and token cost, achieving a 90.15% field-level strict recall on the Spider2-Snow dataset, which could improve the efficiency and accuracy of SQL query generation in complex scenarios.