tubelex-kenlm

TUBELEX Statistical Language Models

N-gram models on the TUBELEX YouTube subtitle corpora. We provide modified Kneser-Ney language models of order 5 (Heafield et al., 2013), i.e. KenLM models.

The files are in LZMA-compressed ARPA format.

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]}
}

• fastText word embeddings
• word frequencies and code

Usage

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

pip install huggingface_hub
pip install https://github.com/kpu/kenlm/archive/master.zip

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

import kenlm
from huggingface_hub import hf_hub_download

model_file = hf_hub_download(repo_id='naist-nlp/tubelex-kenlm', filename='tubelex-en.arpa.xz')
# Loading the model requires KenLM to be compiled with LZMA support (`HAVE_XZLIB`).
# Otherwise you fill first need to decompress the model.
model = kenlm.Model(model_file)

text = ''a sequence of words'  # pre-tokenized, lower-cased, without punctuation
model.perplexity(text)