Skip to content

experiments.md

Latest commit

 

History

History
52 lines (34 loc) · 1.83 KB

experiments.md

File metadata and controls

52 lines (34 loc) · 1.83 KB
Raw

Hyperbolic skip-gram experiments

The python subfolder contains scripts to train and compare skip-gram word embedding models in Euclidean and hyperbolic space. For Euclidean training, the standard implementation in fastText is used.

In order to run the experiments the following variables have to be set in the scripts:

train_skipgram.py

  • INPUT_FILE should point to a preprocessed text file containing one document per line. For our experiments, a 2013 dump of Wikipedia was used. Preprocessing included lower-casing, removal of punctuation and removing articles with less than 20 page views.
  • FASTTEXT_DIR should point to the build directory of fastText's master branch.
  • MINKOWSKI_DIR should point to the build directroy of this project.

Training is then run for a combination of learning rates and dimensions and the resulting word vectors stored in .vec files in the same directory.

evaluate_similarity.py

  • SIMILARITY_DIR is a directory that holds 13 similarity datasets that can be obtained from here.

evaluate_analogy.py

  • ANALOGY_QUESTIONS_FILE is the questions file from the original word2vec repository, available here.

Python dependencies

Please make sure that the hyperboloid_helpers module is on the PYTHONPATH. To install other dependencies run

pip3 install -r requirements.txt

from the python subdirectory.

Usage

The scripts can be run as follows

python3 train_skipgram.py
python3 evaluate_similarity.py
python3 evaluate_analogy.py

These produce printed output on the command line as well as two files results_similarity.txt and results_analogy.txt that save the scores per embedding file.