metadata

annotations_creators:
  - found
language:
  - lv
language_creators:
  - found
license:
  - cc-by-4.0
multilinguality:
  - monolingual
pretty_name: Latvian text dataset
size_categories:
  - 10K<n<100K
source_datasets:
  - extended|tilde_model
  - extended|wikipedia
  - extended|europarl_bilingual
tags:
  - lv
  - latvian
task_categories:
  - automatic-speech-recognition
task_ids: []

Latvian text dataset

Data set of latvian language texts. Intended for use in AI tool development, like speech recognition or spellcheckers

Data sources used

Latvian Wikisource articles - https://wikisource.org/wiki/Category:Latvian
Literary works of Rainis - https://repository.clarin.lv/repository/xmlui/handle/20.500.12574/41
Latvian Wikipedia articles - https://huggingface.co/datasets/joelito/EU_Wikipedias
European Parliament Proceedings Parallel Corpus - https://huggingface.co/datasets/europarl_bilingual
Tilde MODEL Corpus - Multilingual Open Data for European Languages - https://huggingface.co/datasets/tilde_model

To get Wikipedia dataset (197MB) run.

python tools/wikipedia/GetWikipedia.py

To get Europarl dataset (1.7GB) run.

python tools/europarl/GetEuroparl.py

To get Tilde dataset (834MB) run.

python tools/europarl/GetTilde.py

To combine all datasets run

sh combine-all.sh

To clean out some junk run.

sh clean.sh

Also maybe you want to remove duplocate lines. To do so run

sort lv.txt | uniq > lv-uniq.txt

Notes

Possible future sources

Parliament proceedings transcripts - https://www.saeima.lv/lv/transcripts
Discussions of Latvian Wikipedia pages - https://lv.wikipedia.org/wiki/Special:AllPages
Out of copyright books from LNB collection - https://data.gov.lv/dati/lv/dataset/gramatu-digitala-kolekcija

Data sets not used

Web scrapes, as they tend to yield data from comments with improper spelling like "atrashanaas vieta" instead of "atrašanās vieta"
Open Subtitles, as they contain data with improper spelling like "atrashanaas vieta" instead of "atrašanās vieta"

Possible issues:

Data sets contain foreign language characters, like "蠻子" or cyrilic f.e. "Рига"