Repository
BIRD · Text-to-SQL experiments
Purpose
A question in everyday language goes in, a SQL query comes out. One approach breaks the question into smaller decisions, from identifying the entities it mentions to finding the right tables and values. Other experiments retrieve example queries or flag questions that need clarification.
On the financial database in the BIRD benchmark, all 106 queries from the first approach ran successfully. Comparing their answers with the reference results gave 79 matches.