This is the official implementation of the paper "RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL" (AAAI 2023).
If this repository could help you, please cite the following paper:
@inproceedings{li2022resdsql,
author = {Haoyang Li and Jing Zhang and Cuiping Li and Hong Chen},
title = "RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL",
booktitle = "AAAI",
year = "2023"
}
Update (2023.3.13): We evaluated our method on a diagnostic evaluation benchmark, Dr.Spider, which contains 17 test sets to measure the robustness of Text-to-SQL parsers under different perturbation perspectives.
Update (2023.5.19): We added support for CSpider, a Chinese Text-to-SQL benchmark with Chinese questions, English database schema, and corresponding SQL queries.
Update (2023.8.28): Please check out our recent work CodeS, a series of Code LLMs (CodeS-1B, CodeS-3B, CodeS-7B, and CodeS-15B) specifically optimized for SQL generation. You can choose the model that best suits your computational resources and application needs to develop your Text-to-SQL parser!!
We introduce a new Text-to-SQL parser, RESDSQL (Ranking-enhanced Encoding plus a Skeleton-aware Decoding framework for Text-to-SQL), which attempts to decoulpe the schema linking and the skeleton parsing to reduce the difficuty of Text-to-SQL. More details can be found in our paper. All experiments are conducted on a single NVIDIA A100 80G GPU.
We evaluate RESDSQL on six benchmarks: Spider, Spider-DK, Spider-Syn, Spider-Realistic, Dr.Spider, and CSpider. We adopt two metrics: Exact-set-Match accuracy (EM) and EXecution accuracy (EX). Let's look at the following numbers:
On Spider:
| Model | Dev EM | Dev EX | Test EM | Test EX |
|---|---|---|---|---|
| RESDSQL-3B+NatSQL | 80.5% | 84.1% | 72.0% | 79.9% |
| RESDSQL-3B | 78.0% | 81.8% | - | - |
| RESDSQL-Large+NatSQL | 76.7% | 81.9% | - | - |
| RESDSQL-Large | 75.8% | 80.1% | - | - |
| RESDSQL-Base+NatSQL | 74.1% | 80.2% | - | - |
| RESDSQL-Base | 71.7% | 77.9% | - | - |
On Spider-DK, Spider-Syn, and Spider-Realistic:
| Model | DK EM | DK EX | Syn EM | Syn EX | Realistic EM | Realistic EX |
|---|---|---|---|---|---|---|
| RESDSQL-3B+NatSQL | 53.3% | 66.0% | 69.1% | 76.9% | 77.4% | 81.9% |
On Dr.Spider's perturbation sets: Following Dr.Spider, we only report EX for each post-perturbation set and choose PICARD and CodeX as our baseline methods.
| Perturbation set | PICARD | CodeX | RESDSQL-3B | RESDSQL-3B+NatSQL |
|---|---|---|---|---|
| DB-Schema-synonym | 56.5% | 62.0% | 63.3% | 68.3% |
| DB-Schema-abbreviation | 64.7% | 68.6% | 64.5% | 70.0% |
| DB-DBcontent-equivalence | 43.7% | 51.6% | 40.3% | 40.1% |
| NLQ-Keyword-synonym | 66.3% | 55.5% | 67.5% | 72.4% |
| NLQ-Keyword-carrier | 82.7% | 85.2% | 86.7% | 83.5% |
| NLQ-Column-synonym | 57.2% | 54.7% | 57.4% | 63.1% |
| NLQ-Column-carrier | 64.9% | 51.1% | 69.9% | 63.9% |
| NLQ-Column-attribute | 56.3% | 46.2% | 58.8% | 71.4% |
| NLQ-Column-value | 69.4% | 71.4% | 73.4% | 76.6% |
| NLQ-Value-synonym | 53.0% | 59.9% | 53.8% | 53.2% |
| NLQ-Multitype | 57.1% | 53.7% | 60.1% | 60.7% |
| NLQ-Others | 78.3% | 69.7% | 77.3% | 79.0% |
| SQL-Comparison | 68.0% | 66.9% | 70.2% | 82.0% |
| SQL-Sort-order | 74.5% | 57.8% | 79.7% | 85.4% |
| SQL-NonDB-number | 77.1% | 89.3% | 83.2% | 85.5% |
| SQL-DB-text | 65.1% | 72.4% | 67.8% | 74.3% |
| SQL-DB-number | 85.1% | 79.3% | 85.4% | 88.8% |
| Average | 65.9% | 64.4% | 68.2% | 71.7% |
Notice: We also employed the modified test suite script (see this issue) to evaluate the model-generated results, but obtained the same numbers as above. Nevertheless, we suggest that further work should use their modified script to evaluate Dr.Spider.
On CSpider's development set:
| Model | EM | EXEC |
|---|---|---|
| RESDSQL-3B+NatSQL | 66.3% | 81.1% |
| RESDSQL-Large+NatSQL | 64.3% | 81.1% |
| LGESQL + GTL + Electra + QT | 64.0% | - |
| LGESQL + ELECTRA + QT | 64.5% | - |
| RESDSQL-Base+NatSQL | 61.7% | 78.1% |
Create a virtual anaconda environment:
conda create -n your_env_name python=3.8.5Active it and install the cuda version Pytorch:
conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorchInstall other required modules and tools:
pip install -r requirements.txt
pip install https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-2.2.0/en_core_web_sm-2.2.0.tar.gz
python nltk_downloader.pyCreate several folders:
mkdir eval_results
mkdir models
mkdir tensorboard_log
mkdir third_party
mkdir predictionsClone evaluation scripts:
cd third_party
git clone https://github.com/ElementAI/spider.git
git clone https://github.com/ElementAI/test-suite-sql-eval.git
mv ./test-suite-sql-eval ./test_suite
cd ..Download data (including Spider, Spider-DK, Spider-Syn, Spider-Realistic, Dr.Spider, and CSpider) and database and then unzip them:
unzip data.zip
unzip database.zipNotice: Dr.Spider has been preprocessed following the instructions on its Github page.
All evaluation results can be easily reproduced through our released scripts and checkpionts.
Because RESDSQL is a two-stage algorithm, therefore, you should first download cross-encoder checkpoints. Here are links:
| Cross-encoder Checkpoints | Google Drive | Baidu Netdisk |
|---|---|---|
| text2natsql_schema_item_classifier | Link | Link (pwd: 18w8) |
| text2sql_schema_item_classifier | Link | Link (pwd: dr62) |
| xlm_roberta_text2natsql_schema_item_classifier (trained on CSpider) | - | Link (pwd: 3sdu) |
Then, you should download T5 (for Spider) or mT5 (for CSpider) checkpoints:
| T5/mT5 Checkpoints | Google Drive/OneDrive | Baidu Netdisk |
|---|---|---|
| text2natsql-t5-3b | OneDrive link | Link (pwd: 4r98) |
| text2sql-t5-3b | Google Drive link | Link (pwd: sc62) |
| text2natsql-t5-large | Google Drive link | Link (pwd: 7iyq) |
| text2sql-t5-large | Google Drive link | Link (pwd: q58k) |
| text2natsql-t5-base | Google Drive link | Link (pwd: pyxf) |
| text2sql-t5-base | Google Drive link | Link (pwd: wuek) |
| text2natsql-mt5-xl-cspider (trained on CSpider) | - | Link (pwd: y7ei) |
| text2natsql-mt5-large-cspider (trained on CSpider) | - | Link (pwd: ydqk) |
| text2natsql-mt5-base-cspider (trained on CSpider) | - | Link (pwd: d8b8) |
The checkpoints should be placed in the models folder.
For CSpider, we only provide the NatSQL version because its performance is better than SQL in our pre-experiments. To support CSpider, we replace roberta-large with xlm-roberta-large in the first stage and replace t5 with mt5 in the second stage.
The inference scripts are located in scripts/inference.
Concretely, infer_text2natsql.sh is the inference script of RESDSQL-{Base, Large, 3B}+NatSQL, and infer_text2sql.sh is the inference script of RESDSQL-{Base, Large, 3B}. For example, you can run the inference of RESDSQL-3B+NatSQL on Spider's dev set via:
sh scripts/inference/infer_text2natsql.sh 3b spiderThe first argument (model scale) can be selected from [base, large, 3b] and the second argument (dataset name) can be selected from [spider, spider-realistic, spider-syn, spider-dk, DB_schema_synonym, DB_schema_abbreviation, DB_DBcontent_equivalence, NLQ_keyword_synonym, NLQ_keyword_carrier, NLQ_column_synonym, NLQ_column_carrier, NLQ_column_attribute, NLQ_column_value, NLQ_value_synonym, NLQ_multitype, NLQ_others, SQL_comparison, SQL_sort_order, SQL_NonDB_number, SQL_DB_text, SQL_DB_number].
The predicted SQL queries are recorded in predictions/{dataset_name}/{model_name}/pred.sql.
Inference on CSpider's Dev Set (New Feature) We also provide inference scripts to run RESDSQL-{Base, Large, 3B}+NatSQL on CSpider's development set. Here is an example:
sh scripts/inference/infer_text2natsql_cspider.sh 3bThe first argument (model scale) can be selected from [base, large, 3b].
We provide scripts in scripts/train/text2natsql and scripts/train/text2sql to train RESDSQL on Spider's training set and evaluate on Spider's dev set.
RESDSQL-{Base, Large, 3B}+NatSQL
# Step1: preprocess dataset
sh scripts/train/text2natsql/preprocess.sh
# Step2: train cross-encoder
sh scripts/train/text2natsql/train_text2natsql_schema_item_classifier.sh
# Step3: prepare text-to-natsql training and development set for T5
sh scripts/train/text2natsql/generate_text2natsql_dataset.sh
# Step4: fine-tune T5-3B (RESDSQL-3B+NatSQL)
sh scripts/train/text2natsql/train_text2natsql_t5_3b.sh
# Step4: (or) fine-tune T5-Large (RESDSQL-Large+NatSQL)
sh scripts/train/text2natsql/train_text2natsql_t5_large.sh
# Step4: (or) fine-tune T5-Base (RESDSQL-Base+NatSQL)
sh scripts/train/text2natsql/train_text2natsql_t5_base.shRESDSQL-{Base, Large, 3B}
# Step1: preprocess dataset
sh scripts/train/text2sql/preprocess.sh
# Step2: train cross-encoder
sh scripts/train/text2sql/train_text2sql_schema_item_classifier.sh
# Step3: prepare text-to-sql training and development set for T5
sh scripts/train/text2sql/generate_text2sql_dataset.sh
# Step4: fine-tune T5-3B (RESDSQL-3B)
sh scripts/train/text2sql/train_text2sql_t5_3b.sh
# Step4: (or) fine-tune T5-Large (RESDSQL-Large)
sh scripts/train/text2sql/train_text2sql_t5_large.sh
# Step4: (or) fine-tune T5-Base (RESDSQL-Base)
sh scripts/train/text2sql/train_text2sql_t5_base.shDuring training, the cross-encoder (i.e., the first stage) always keeps the best checkpoint, but T5 (i.e., the second stage) keeps all the intermediate checkpoints, because different test sets may achieve the best Text-to-SQL performance on different checkpoints. Therefore, given a test set, we need to evaluate all the intermediate checkpoints and compare their performance to find the best checkpoint. The evaluation results of checkpoints are saved in eval_results.
Our paper also report the performence of RESDSQL-3B+NatSQL (the most powerful version of RESDSQL) on Spider-DK, Spider-Syn, and Spider-Realistic. To obtain results on these datasets, we provide evaluation scripts in scripts/evaluate_robustness. Here is an example for Spider-DK:
# Step1: preprocess Spider-DK
sh scripts/evaluate_robustness/preprocess_spider_dk.sh
# Step2: Run evaluation on Spider-DK
sh scripts/evaluate_robustness/evaluate_on_spider_dk.shWe additionally provide scripts in scripts/train/cspider_text2natsql and scripts/train/cspider_text2sql to train RESDSQL on CSpider's training set and evaluate on CSpider's dev set.
RESDSQL-{Base, Large, 3B}+NatSQL (CSpider version)
# Step1: preprocess CSpider
sh scripts/train/cspider_text2natsql/preprocess.sh
# Step2: train cross-encoder
sh scripts/train/cspider_text2natsql/train_text2natsql_schema_item_classifier.sh
# Step3: prepare text-to-natsql training and development set for mT5
sh scripts/train/cspider_text2natsql/generate_text2natsql_dataset.sh
# Step4: fine-tune mT5-XL (RESDSQL-3B+NatSQL)
sh scripts/train/cspider_text2natsql/train_text2natsql_mt5_xl.sh
# Step4: (or) fine-tune mT5-Large (RESDSQL-Large+NatSQL)
sh scripts/train/cspider_text2natsql/train_text2natsql_mt5_large.sh
# Step4: (or) fine-tune mT5-Base (RESDSQL-Base+NatSQL)
sh scripts/train/cspider_text2natsql/train_text2natsql_mt5_base.shIn order to train the NatSQL version on CSpider, we manually aligned and modified annotations of NatSQL. The aligned files are also released, see NatSQL/NatSQLv1_6/train_cspider-natsql.json and NatSQL/NatSQLv1_6/dev_cspider-natsql.json.
RESDSQL-{Base, Large, 3B} (CSpider version)
# Step1: preprocess CSpider
sh scripts/train/cspider_text2sql/preprocess.sh
# Step2: train cross-encoder
sh scripts/train/cspider_text2sql/train_text2sql_schema_item_classifier.sh
# Step3: prepare text-to-sql training and development set for mT5
sh scripts/train/cspider_text2sql/generate_text2sql_dataset.sh
# Step4: fine-tune mT5-XL (RESDSQL-3B)
sh scripts/train/cspider_text2sql/train_text2sql_mt5_xl.sh
# Step4: (or) fine-tune mT5-Large (RESDSQL-Large)
sh scripts/train/cspider_text2sql/train_text2sql_mt5_large.sh
# Step4: (or) fine-tune mT5-Base (RESDSQL-Base)
sh scripts/train/cspider_text2sql/train_text2sql_mt5_base.shWe would thanks to Hongjin Su and Tao Yu for their help in evaluating our method on Spider's test set. We would also thanks to PICARD (paper, code), NatSQL (paper, code), Spider (paper, dataset), Spider-DK (paper, dataset), Spider-Syn (paper, dataset), Spider-Realistic (paper, dataset), Dr.Spider (paper, dataset), and CSpider (paper, dataset) for their interesting work and open-sourced code and dataset.