[Improvement] Advanced SQL Generation (RESDSQL)

Open alex-sun-kuo opened this issue 2 years ago • 1 comments

https://github.com/RUCKBReasoning/RESDSQL

^ This repo has an implementation of this paper: https://arxiv.org/pdf/2302.05965.pdf It details an advanced strategy for generating accurate SQL to answer a natural language query. As far as I understand there's a muli-step sequential process:

table selection
column selection
generation of a SQL skeleton (think madlib for SQL queries lol)
filling in that SQL skeleton

It would be dope to implement this in our project. It would help us scale up to larger datasets.

Currently, we just do:

table selection
generate SQL given these tables and their columns

Screenshot 2023-03-13 at 4 51 06 PM

Mar 13 '23 23:03 alex-sun-kuo

wow interesting stuff. this is optimizing on the encoder / decoder level for optimal SQL generation.

Ranking-enhanced Encoding and Skeleton-aware Decoding

I'm gonna try to understand this. Also tried looking for this model on HuggingFace but couldn't find it, otherwise that would be pretty awesome

Mar 14 '23 02:03 asaprahul