/text2sql-lgesql

This is the project containing source codes and pre-trained models about ACL2021 Long Paper ``LGESQL: Line Graph Enhanced Text-to-SQL Model with Mixed Local and Non-Local Relations".

Primary LanguagePythonApache License 2.0Apache-2.0

LGESQL

This is the project containing source code for the paper LGESQL: Line Graph Enhanced Text-to-SQL Model with Mixed Local and Non-Local Relations in ACL 2021 main conference. If you find it useful, please cite our work.

    @inproceedings{cao-etal-2021-lgesql,
            title = "{LGESQL}: Line Graph Enhanced Text-to-{SQL} Model with Mixed Local and Non-Local Relations",
            author = "Cao, Ruisheng  and
            Chen, Lu  and
            Chen, Zhi  and
            Zhao, Yanbin  and
            Zhu, Su  and
            Yu, Kai",
            booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
            month = aug,
            year = "2021",
            address = "Online",
            publisher = "Association for Computational Linguistics",
            url = "https://aclanthology.org/2021.acl-long.198",
            doi = "10.18653/v1/2021.acl-long.198",
            pages = "2541--2555",
    }

Create environment and download dependencies

The following commands are provided in setup.sh.

  1. Firstly, create conda environment text2sql:
  • In our experiments, we use torch==1.6.0 and dgl==0.5.3 with CUDA version 10.1

  • We use one GeForce RTX 2080 Ti for GLOVE and base-series pre-trained language model~(PLM) experiments, one Tesla V100-PCIE-32GB for large-series PLM experiments

    conda create -n text2sql python=3.6
    source activate text2sql
    pip install torch==1.6.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html
    pip install -r requirements.txt
    
  1. Next, download dependencies:

     python -c "import stanza; stanza.download('en')"
     python -c "from embeddings import GloveEmbedding; emb = GloveEmbedding('common_crawl_48', d_emb=300)"
     python -c "import nltk; nltk.download('stopwords')"
    
  2. Download pre-trained language models from Hugging Face Model Hub, such as bert-large-whole-word-masking and electra-large-discriminator, into the pretrained_models directory. The vocab file for glove.42B.300d is also pulled: (please ensure that Git LFS is installed)

     mkdir -p pretrained_models && cd pretrained_models
     git lfs install
     git clone https://huggingface.co/bert-large-uncased-whole-word-masking
     git clone https://huggingface.co/google/electra-large-discriminator
     mkdir -p glove.42b.300d && cd glove.42b.300d
     wget -c http://nlp.stanford.edu/data/glove.42B.300d.zip && unzip glove.42B.300d.zip
     awk -v FS=' ' '{print $1}' glove.42B.300d.txt > vocab_glove.txt
    

Download and preprocess dataset

  1. Download, unzip and rename the spider.zip into the directory data.

  2. Merge the data/train_spider.json and data/train_others.json into one single dataset data/train.json.

  3. Preprocess the train and dev dataset, including input normalization, schema linking, graph construction and output actions generation. (Our preprocessed dataset can be downloaded here)

     ./run/run_preprocessing.sh
    

Training

Training LGESQL models with GLOVE, BERT and ELECTRA respectively:

  • msde: mixed static and dynamic embeddings

  • mmc: multi-head multi-view concatenation

    ./run/run_lgesql_glove.sh [mmc|msde]
    ./run/run_lgesql_plm.sh [mmc|msde] bert-large-uncased-whole-word-masking
    ./run/run_lgesql_plm.sh [mmc|msde] electra-large-discriminator
    

Evaluation and submission

  1. Create the directory saved_models, save the trained model and its configuration (at least containing model.bin and params.json) into a new directory under saved_models, e.g. saved_models/electra-msde-75.1/.

  2. For evaluation, see run/run_evaluation.sh and run/run_submission.sh (eval from scratch) for reference.

  3. Model instances and submission scripts are available in codalab:plm and google drive: including submitted BERT and ELECTRA models. Codes and model for GLOVE are deprecated.

Results

Dev and test EXACT MATCH ACC in the official leaderboard, also provided in the results directory:

model dev acc test acc
LGESQL + GLOVE 67.6 62.8
LGESQL + BERT 74.1 68.3
LGESQL + ELECTRA 75.1 72.0

Acknowledgements

We would like to thank Tao Yu, Yusen Zhang and Bo Pang for running evaluations on our submitted models. We are also grateful to the flexible semantic parser TranX that inspires our works.