Awesome

XCFGs

Aiming at unifying all extensions of context-free grammars (XCFGs). X stands for weighted, (compound) probabilistic, and neural extensions, etc. Currently only the data preprocessing module has been implemented though.

Update (08/06/2023): Support Brown Corpus and English Web Treebank that are used in this study.

Update (06/02/2022): Parse MSCOCO and Flickr30k captions, create data splits, and encode images for VC-PCFG.

Update (03/10/2021): Parallel Chinese-English data is supported.

Data

The repo handles WSJ, CTB, SPMRL, Brown Corpus, and English Web Treebank. Have a look at treebank.py.

If you are looking for the data used in C-PCFGs. Follow the instructions in treebank.py and put all outputs in the same folder, let us say ./data.punct. The script only removes morphology features and creates data splits. To remove punctuation we will need clean_tb.py. For example, I used python clean_tb.py ./data.punct ./data.clean. All the cleaned treebanks will reside in /data.clean. Then simply execute the command ./batchify.sh ./data.clean/, you will have all the data needed to reproduce the results in C-PCFGs. Feel free to change parameters in batchify.sh if you want to use a different batch size or vocabulary size.

Evaluation

To ease evaluation I represent a gold tree as a tuple:

TREE: TUPLE(sentence: STR, spans: LIST[SPAN], span_labels: LIST[STR], pos_tags: LIST[STR])
SPAN: TUPLE(left_boundary: INT, right_boundary: INT)

If you have followed the instructions in the last section, this command ./binarize.sh ./data.clean/ could help you convert gold trees into the tuple representation.

Trivial baselines

Even for trivial baselines, e.g., left- and right-branching trees, you may find different F1 numbers in literature on grammar induction, partly because the authors used (slightly) different procedures for data preprocessing. To encourage truly fair comparison I also released a standard procedure baseline.py. Hopefully, this will help with the situation.

An evaluation checklist for phrase-structure grammar induction

Below is a comparison of several cirtical training / evaluation settings of recent unsupervised parsing models.

Citing XCFGs

If you use XCFGs in your research or wish to refer to the results in C-PCFGs, please use the following BibTeX entry.

@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",
}

Acknowledgements

batchify.py is borrowed from C-PCFGs.

License

MIT