Fast and Modularized Implementations of XCFGs

Update (08/06/2023): Add an evaluation of the transferability of VC-PCFG.

Update (12/12/2021): Add an implementation of VC-PCFG.

Update (03/10/2021): Fast and modularized implementations of XCFGs.

Model

This implementation reaches an average sentence-level F1 56%, slightly higher than Yoon's 55.2%. Notably, it takes only 25 minutes per epoch on a GeForce GTX 1080 Ti. Here is the report: An Empirical Study of Compound PCFGs.

Data

I am using the same data processing as Yoon. If you are looking for a unified data pipeline for WSJ, CTB, and SPMRL, I suggest you take a look at XCFG. It makes data creation easier. If you still find it annoying processing all the data from scratch, contact me and I can give you access to all the processed data (please make sure you have acquired licenses for these treebanks).

Mean sentence-level F1 numbers

Here is an overview of model performance on WSJ, CTB, and SPMRL. Find more details in the report.

On WSJ and CTB

Model	WSJ	CTB
Yoon's	55.2	36.0
This Repo	55.7_±1.3	35.1_±6.1

On SPMRL

Model	Basque	German	French	Hebrew	Hungarian	Korean	Polish	Swedish
N-PCFG	30.2_±0.9	37.8_±1.7	42.2_±1.4	41.0_±0.6	37.9_±0.8	25.7_±2.8	31.7_±1.8	14.5_±12.7
C-PCFG	27.9_±2.0	37.3_±1.8	40.5_±0.8	39.2_±1.2	38.3_±0.7	27.7_±2.8	32.4_±1.1	23.7_±14.3

Learning

Specify the model saving path M_ROOT and the data path D_ROOT before running. Check out XCFG if you are unsure about how to prepare data.

python train.py num_gpus=1 eval=False alias_root=$M_ROOT data.data_root=$D_ROOT \
    running.peep_rate=500 running.save_rate=1e9 running.save_epoch=True data.eval_samples=50000 \
    +model/pcfg=default \
    +optimizer=default \
    +data=default \
    +running=default

Parsing

Inference needs two more runninng parameters than learning: (1) M_NAME is the name of an experiment you have run and want to test and (2) M_FILE is the name of a model weight file from the experiment you have run.

python train.py num_gpus=1 eval=True alias_root=$M_ROOT data.data_root=$D_ROOT \
    model_name=$M_NAME model_file=$M_FILE data.eval_samples=50000 data.eval_name=english-test.json \
    +model/pcfg=default \
    +optimizer=default \
    +data=default \
    +running=default

Dependencies

It requires a tailored Torch-Struct.

git clone --branch beta https://github.com/zhaoyanpeng/cpcfg.git
cd cpcfg
virtualenv -p python3.7 ./pyenv/oops
source ./pyenv/oops/bin/activate
pip install --upgrade pip
pip install -r requirements.txt
git clone --branch infer_pos_tag https://github.com/zhaoyanpeng/pytorch-struct.git
cd pytorch-struct
pip install -e .

Citation

If you use the fast implementation of C-PCFGs in your research or wish to refer to the results in the report, please use the following BibTeX entries.

@inproceedings{zhao-titov-2023-transferability,
    title = "On the Transferability of Visually Grounded {PCFGs}",
    author = "Zhao, Yanpeng  and Titov, Ivan",
    booktitle = "Findings of the Association for Computational Linguistics: EMNLP 2023",
    month = dec,
    year = "2023",
    address = "Singapore",
    publisher = "Association for Computational Linguistics",
}

@inproceedings{zhao-titov-2021-empirical,
    title = "An Empirical Study of Compound {PCFG}s",
    author = "Zhao, Yanpeng and Titov, Ivan",
    booktitle = "Proceedings of the Second Workshop on Domain Adaptation for NLP",
    month = apr,
    year = "2021",
    address = "Kyiv, Ukraine",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.adaptnlp-1.17",
    pages = "166--171",
}

@inproceedings{zhao-titov-2020-visually,
    title = "Visually Grounded Compound {PCFG}s",
    author = "Zhao, Yanpeng  and Titov, Ivan",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2020.emnlp-main.354",
    doi = "10.18653/v1/2020.emnlp-main.354",
    pages = "4369--4379",
}

Acknowledgements

This repo is developed based on C-PCFGs and Torch-Struct.

License

MIT

zhaoyanpeng / cpcfg