tubelex-fasttext

TUBELEX FastText Word Embeddings

FastText Word Embeddings trained on the TUBELEX YouTube subtitle corpora. We use the 300-dimensional fastText CBOW model with position weights, 10 negative samples, 10 epochs, character 5-grams (other paramters: default) (Grave et al., 2018).

We provide both '*.bin' files (for fastText) and '*.vec' files that follow the common Word2vec format, and can be used for instance with the gensim package.

What is TUBELEX?

TUBELEX is a YouTube subtitle corpus currently available for Chinese, English, Indonesian, Japanese, and Spanish.

• preprint, BibTeX entry:

@article{nohejl_etal_2024_film,
  title={Beyond {{Film Subtitles}}: {{Is YouTube}} the {{Best Approximation}} of {{Spoken Vocabulary}}?},
  author={Nohejl, Adam and Hudi, Frederikus and Kardinata, Eunike Andriani and Ozaki, Shintaro and Riera Machin, Maria Angelica and Sun, Hongyu and Vasselli, Justin and Watanabe, Taro},
  year={2024}, eprint={2410.03240}, archiveprefix={arXiv}, primaryclass={cs.CL},
  url={https://arxiv.org/abs/2410.03240v1}, journal={ArXiv preprint}, volume={arXiv:2410.03240v1 [cs]}
}

• KenLM n-gram models
• word frequencies and code

Usage

To download and use the fastText models in Python, first install dependencies:

pip install huggingface_hub
pip install fasttext

You can then use e.g. the English (en) model in the following way:

import fasttext
from huggingface_hub import hf_hub_download

model_file = hf_hub_download(repo_id='naist-nlp/tubelex-kenlm', filename='tubelex-en.bin')
model = fasttext.load_model(model_file)

print(model['koala'])