Tensor2Struct

tensor2struct is a package that contains a set of neural semantic parsers based on the encoder-decoder framework. Currently, it supports the following datasets:

• Overnight
• Spider
• Chinese Spider
• SCAN
• COGS

Setup

Create a virtual environment and run the setup script.

conda create --name tensor2struct python=3.7
conda activate tensor2struct
./setup.sh

wandb is used for logging. To enable it you can create your own account and wandb login to enable logging. Or you could just wandb off to only allow for dryrun locally.

In general, the raw data is expected to be placed under "/data/TASK_NAME/raw" where TASK_NAME could be spider/ssp/overnight.

Make log/ and ie_dir/ which will be used for storing checkpoints and predictions (during inference).

Experiments

Tensor2struct has been the backbone architecture for implementing new models, objectives, algorithms proposed in the following papers. To reproduce experiments from a paticular paper, the corresponding link below will take to detailed instructions.

• Meta-Learning for Domain Generalization
• Learning from Executions for Semantic Parsing
• Learning to Synthesize Data for Semantic Parsing
• Meta-Learning to Compositionally Generalize
• Structured Reordering for Modeling Latent Alignments in Sequence Transduction

Citations

If you use tensor2struct, please cite one of the following papers.

• Meta-Learning for Domain Generalization in Semantic Parsing, NAACL 2021

bibtex

@inproceedings{wang-etal-2021-meta,
    title = "Meta-Learning for Domain Generalization in Semantic Parsing",
    author = "Wang, Bailin  and
      Lapata, Mirella  and
      Titov, Ivan",
    booktitle = "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jun,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2021.naacl-main.33",
    doi = "10.18653/v1/2021.naacl-main.33",
    pages = "366--379",
}

• Learning from Executions for Semantic Parsing, NAACL 2021

bibtex

@inproceedings{wang-etal-2021-learning-executions,
    title = "Learning from Executions for Semantic Parsing",
    author = "Wang, Bailin  and
      Lapata, Mirella  and
      Titov, Ivan",
    booktitle = "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jun,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2021.naacl-main.219",
    doi = "10.18653/v1/2021.naacl-main.219",
    pages = "2747--2759",
}

• Learning to Synthesize Data for Semantic Parsing, NAACL 2021

bibtex

@inproceedings{wang-etal-2021-learning-synthesize,
    title = "Learning to Synthesize Data for Semantic Parsing",
    author = "Wang, Bailin  and
      Yin, Wenpeng  and
      Lin, Xi Victoria  and
      Xiong, Caiming",
    booktitle = "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
    month = jun,
    year = "2021",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2021.naacl-main.220",
    doi = "10.18653/v1/2021.naacl-main.220",
    pages = "2760--2766",
}

• Meta-Learning to Compositionally Generalize, ACL 2021

bibtex

@inproceedings{conklin-etal-2021-meta,
    title = "Meta-Learning to Compositionally Generalize",
    author = "Conklin, Henry  and
      Wang, Bailin  and
      Smith, Kenny  and
      Titov, Ivan",
    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.258",
    doi = "10.18653/v1/2021.acl-long.258",
    pages = "3322--3335",
}

• Structured Reordering for Modeling Latent Alignments in Sequence Transduction, NeurIPS 2021

bibtex

@inproceedings{
wang2021structured,
title={Structured Reordering for Modeling Latent Alignments in Sequence Transduction},
author={bailin wang and Mirella Lapata and Ivan Titov},
booktitle={Thirty-Fifth Conference on Neural Information Processing Systems},
year={2021},
url={https://openreview.net/forum?id=X2Cxixkcpx}
}

Acknowledgement

Tensor2struct is a generalization of RAT-SQL.