Efficient Ambiguity Resolution via Information-Guided Clarification for Text-to-SQL
TL;DR
This work presents an interactive Text-to-SQL framework that models SQL generation as probabilistic reasoning over multiple candidate queries and selects clarification questions based on the principle of Expected Information Gain, providing an information-theoretic criterion for prioritizing high-impact clarifications and thus enabling efficient disambiguation and higher accuracy.
Abstract
Text-to-SQL technology enables users to query relational databases using natural language, but ambiguity in user questions remains a major obstacle: non-expert users naturally produce ambiguous queries, yet current Text-to-SQL systems often ignore ambiguity or rely on interaction without a principled strategy, resulting in inaccurate interpretations or excessive interaction cost. We address this challenge by framing ambiguity resolution as an uncertainty reduction problem and focusing on interaction efficiency. Specifically, we present an interactive Text-to-SQL framework that models SQL generation as probabilistic reasoning over multiple candidate queries and selects clarification questions based on the principle of Expected Information Gain (EIG), providing an information-theoretic criterion for prioritizing high-impact clarifications and thus enabling efficient disambiguation and higher accuracy. To support systematic training and evaluation, we also introduce an automated LLM-based pipeline for constructing large-scale ambiguous Text-to-SQL data with ambiguity explanations, avoiding costly manual annotation. Experiments on multiple benchmarks demonstrate that our approach achieves higher accuracy with fewer clarification turns, highlighting the importance of information-guided interaction for efficient ambiguity resolution.