***** New June 28st, 2019 *****
Fix the execution accuracy calculation, covering all correct cases, for example
SELECT COL_1, COL_2 ... SELECT COL_2, COL_1 ...
Another example is
SELECT COL_1 ... WHERE COL_2 == 1 OR COL_3 == 1 SELECT COL_1 ... WHERE COL_3 == 1 OR COL_2 == 1
Now we assume these SQL pairs have the same execution result.
***** New June 18st, 2019 *****
This version of release supports execution accuracy, which gets the execution result of predicted SQL. This requires records==0.5.3 before running.
The model decouples the task of generating a whole SQL into several sub-tasks, including select-number, select-column, select-aggregation, condition-number, condition-column and so on.
Simple model structure shows here, implementation details could refer to the origin paper.
The difference between SQLNet and this baseline model is, Select-Number and Where-Relationship sub-tasks are added to adapt this Chinese NL2SQL dataset better.
- Python 2.7
- torch 1.0.1
- records 0.5.3
Start to train
Firstly, download the provided datasets at ~/data_nl2sql/, which should include train.json, train.tables.json, val.json, val.tables.json and char_embedding, and divide them in following structure.
├── data_nl2sql │ ├── train │ │ ├── train.db │ │ ├── train.json │ │ ├── train.tables.json │ ├── val │ │ ├── val.db │ │ ├── val.json │ │ ├── val.tables.json │ ├── test │ │ ├── test.db │ │ ├── test.json │ │ ├── test.tables.json │ ├── char_embedding
mkdir ~/nl2sql cd ~/nl2sql/ git clone https://github.com/ZhuiyiTechnology/nl2sql_baseline.git cp -r ~/data_nl2sql/* ~/nl2sql/nl2sql_baseline/data/ cd ~/nl2sql/nl2sql_baseline/ sh ./start_train.sh 0 128
while the first parameter 0 means gpu number, the second parameter means batch size.
Start to evaluate
To evaluate on val.json or test.json, make sure trained model is ready, then run
cd ~/nl2sql/nl2sql_baseline/ sh ./start_test.sh 0 pred_example
while the first parameter 0 means gpu number, the second parameter means the output path of prediction.
We have run experiments several times, achiving avegrage 27.5% logic form accuracy on the val dataset, with 128 batch size.
We found the main challenges of this datasets containing poor condition value prediction, select column and condition column not mentioned in NL question, inconsistent condition relationship representation between NL question and SQL, etc. All these challenges could not be solved by existing baseline and SOTA models.
Correspondingly, this baseline model achieves only 77% accuracy on condition column and 62% accuracy on condition value respectively even on the training set, and the overall logic form is only around 50% as well, indicating these problems are challenging for contestants to solve.