This repository contains the code for Word2Vec with Structure Prediction and Word2Vec with Structure Constraint from the paper Domain-Specific Word Embeddings with Structure Prediction (to appear in TACL).

@article{brandl2022domain,
  title={Domain-Specific Word Embeddings with Structure Prediction},
  author={Brandl, Stephanie and Lassner, David and Baillot, Anne and Nakajima, Shinichi},
  journal={arXiv preprint arXiv:2210.04962},
  year={2022}
}

After cloning the repository, we recommend to set up a virtual environment where you can install the code via pip install -e ., then all required packages will be downloaded.

With python run.py you should be able to download the dataset Wikipedia Field of Science and compute domain-specific word embeddings, you can choose between:

model = W2VPred(tau=1024, lam=512, V=len(vocab), T=len(slices), d=d)

and

w = tree_distance_matrix_to_w(get_wikipedia_fos_gt())  
model = W2VConstr(tau=256, lam=512, V=len(vocab), T=len(slices), d=d, w=w)

uncomment accordingly in run.py and set your paths in src.domain_word_mbeddings.configuration.py

In case of questions or concerns, you can reach us via brandl[at]di.ku.dk and lassner[at]tu-berlin.de.