AntoineSimoulin / opus-open-subtitles-monolingual

Preprocessing of the opus open subtitles corpus

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

opus-open-subtitles-monolingual

Contain scripts to preprocess the Open Subtitle mono lingual corpus. To avoid to many I/O on disk, subtitles documents are handled in queues. You first load queue_len document in memory. You can then process them (parse XML and de-tokenize) using multiprocessing. You can set the number of workers with n_procs.

You can download files from the opus website:

wget https://opus.nlpl.eu/download.php?f=OpenSubtitles/v2018/xml/fr.zip -O fr.zip

You can then preprocess the file into jsonl format with one line per subtitle:

python preprocess.py --input_dir /data/.. --output_dir /data/... --queue_len 10000 --n_procs 5 --language fr

Finally you ay compress the file to optimize disk space:

gzip -c fr.jsonl > fr.jsonl.gz

You may then read the dataset with HuggingFace datasetslibrary:

from datasets import load_dataset

dataset = load_dataset(
  'json', 
  data_files='{}.jsonl'.format("fr"))

About

Preprocessing of the opus open subtitles corpus


Languages

Language:Python 100.0%