Skip to content
main
Switch branches/tags
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
img
 
 
 
 
 
 
 
 
 
 
 
 

ComHyper License: MIT

Code for EMNLP'20 paper "When Hearst Is not Enough: Improving Hypernymy Detection from Corpus with Distributional Models" (arXiv)

In a nutshell, ComHyper is the complementary framework for solving hypernymy detection tasks from the perspective of blind points of Hearst pattern-based methods. As shown in the left Figure, long-tailed nouns cannot well covered by Hearst patterns and thus form non-negligible sparsity types. For such cases, we propose to use supervised distributional models for complmenting pattern-based models shown in the right Figure.

Use ComHyper

1. Download Hearst pattern files and corpus.

First prepare the extracted Hearst pattern pairs such as hearst_counts.txt.gz from the repo hypernymysuite or data-concept.zip from Microsoft Concept Graph (Also known as Probase). Specify the parameter pattern_filename in the config as the file location.

wget https://github.com/facebookresearch/hypernymysuite/blob/master/hearst_counts.txt.gz
curl -L "https://concept.research.microsoft.com/Home/StartDownload" > data-concept.zip

Then extract the contexts for words from large-scale corpus such as Wiki + Gigaword or ukWac. All the contexts for one word should be organized into one txt file and one line for one context.

For those words appearing in the Hearst patterns (IP words), organize their context files into the directory context in the config. For OOP words, organize their context files into the context_oov in the config.

2. Train and evaluate the ComHyper.

For training the distributional models supervsied by the output of pattern-based models, different context encoders are provided:

python train_word2score.py config/word.cfg
python train_context2score.py config/context.cfg
python train_bert2score.py config/bert.cfg

The same evaluation scripts work for all settings. For reproducing the results, run:

python evaluation/evaluation_all_context.py ../config/context.cfg 

Note that we choose not to report the BERT encoder results in our orginial paper due to efficiency but release the relevant codes for incoroporating effective pre-trained contextualized encoders to further improve the performance. Welcome to PR or contact cyuaq # cse.ust.hk !

Citation

Please cite the following paper if you found our method helpful. Thanks !

@inproceedings{yu-etal-2020-hearst,
    title = "When Hearst Is Not Enough: Improving Hypernymy Detection from Corpus with Distributional Models",
    author = "Yu, Changlong and Han, Jialong and Wang, Peifeng and Song, Yangqiu and Zhang, Hongming and Ng, Wilfred and Shi, Shuming",
    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://www.aclweb.org/anthology/2020.emnlp-main.502",
    pages = "6208--6217",
}

About

Code for EMNLP'20 paper "When Hearst Is not Enough: Improving Hypernymy Detection from Corpus with Distributional Models"

Topics

Resources

License

Releases

No releases published

Packages

No packages published