A sample t5-small finetuned on wikiSQL. The method used is fairly simple, we treat the problem as seq2Seq translation from natural language to an intermediate language, in the spirit of SeeD (https://arxiv.org/abs/2105.07911). The project follows standard datascience cookiecutter naming patterns. Make targets are as follows.
make data_finetune
make train_finetune
make predict_finetune
The output format is a small custom language I've created to succinctly express SQL queries. A couple of examples from the validation dataset are:
Natural language question: [START]-col0:District, col1:Total amount of trees, col2:Prevailing types, %, col3:Amount of old trees, col4:Amount of trees, that require replacement, ---What is the district when prevailing types, % is acer negundo — 30.22 tilia — 18.6 poplar — 15.23?-
Actual SQL: [START_SQL] 0:0:ci:[2]::oi:[0]:c['Acer negundo — 30.22 Tilia — 18.6 Poplar — 15.23']::
Predicted SQL: <pad> [START_SQL] 0:0:ci:[2]::oi:[0]:c['Acer Negundo — 30.22 Tilia — 18.6 Poplar — 15.23']::
Natural language question: [START]-col0:Year, col1:Division, col2:League, col3:Regular Season, col4:Playoffs, col5:U.S. Open Cup, ---what's the u.s. open cup status for regular season of 4th, atlantic division -
Actual SQL: [START_SQL] 5:0:ci:[3]::oi:[0]:c['4th, Atlantic Division']::
Predicted SQL: <pad> [START_SQL] 5:0:ci:[3]::oi:[0]:c['4th, Atlantic Division']::
`