This repository contains the code used in the paper "A Systematic Comparison of English Noun Compound Representations" in the MWE workshop @ ACL 2019. It can be used to train various types of noun compound embeddings and evaluate them on several tasks.

• Python 3
• argparse
• allennlp (0.8.2)
• gensim
• word_forms

The vocabulary is based on the Tratz (2011) dataset. It is found under data/nc_vocab.txt. This repository does not handle recognizing noun compounds.

I used the English Wikipedia dump from January 2018 as the corpus. To download:

wget https://dumps.wikimedia.org/enwiki/20180101/enwiki-20180101-pages-meta-current.xml.bz2;

Then, we extract and clean the text using WikiExtractor:

python WikiExtractor.py --processes 20 -o corpora/text/ ~/corpora/[lang]wiki-20180101-pages-meta-current.xml.bz2;
cat corpora/text/*/* > output/en_corpus;

Finally, tokenize the corpus using spacy:

python training/distributional/preprocessing/tokenize_corpus.py output/en_corpus;

You can train embeddings with several training objectives:

• Distributional - learning vectors by treating noun compounds as single tokens. See here.
• Compositional - learning vectors by composing the embeddings of the constituent words. See here.
• Paraphrase Based - learning vectors by minimizing the distance to the embeddings of paraphrases. See here.

The noun compound embeddings were evaluated on the following tasks:

• Top K - a qualitative analysis of the neighbours of each noun compound. See here.
• Classification - multiclass classification of the noun compounds according to the relationship between the head and the modifier. See here.
• Attributes - binary classification to whether a noun compound holds a certain property. See here.

If you used this code in your publication, please cite the following paper:

@inproceedings{shwartz-2019-nc_embeddings,
    title = "A Systematic Comparison of English Noun Compound Representations",
    author = "Shwartz, Vered",
    booktitle = "Proceedings of the Joint Workshop on Multiword Expressions and WordNet (MWE-WN 2019) @ ACL 2019",
    month = August,
    year = "2019",
    address = "Florence, Italy",
    publisher = "Association for Computational Linguistics",
}