Rule-SQL

https://arxiv.org/abs/1907.00620

(Incorporating database design info into weak supervised text-to-sql generation)

REASONABLE

The rules of database itself are filtering the candidate SQLs which lead to the same answer. This method can combine other weakly supervised text2sql methods.

Run

Run data/add_answer.py to add answer for the origin wikisql data.

Run rl/data.py for the SQL exploration.

Data

https://drive.google.com/file/d/1iJvsf38f16el58H4NPINQ7uzal5-V4v4/

Result without EG （1 condition）

Model	Dev logical form accuracy	Dev execution accuracy	Test logical form accuracy	Test execution accuracy
Neural Symbolic Machines	-	74.9	-	74.8
Our Methods	68.7	81.2	68.5	81.0

Reference

https://github.com/naver/sqlova