Datasets:

Modalities:
Text
Formats:
parquet
DOI:
Libraries:
Datasets
pandas
License:
fdelucaf commited on
Commit
d8737de
·
verified ·
1 Parent(s): 1942bca

Improved description in Source Data section

Browse files
Files changed (1) hide show
  1. README.md +6 -3
README.md CHANGED
@@ -69,9 +69,12 @@ This dataset is aimed at promoting the development of Machine Translation betwee
69
  The dataset is a combination of the following original datasets collected from [Opus](https://opus.nlpl.eu/):
70
  NLLB, MultiCCAligned, WikiMatrix, GNOME, KDE4, OpenSubtitles, GlobalVoices.
71
 
72
- All datasets are deduplicated and filtered to remove any sentence pairs with a cosine similarity of less than 0.75.
73
- This is done using sentence embeddings calculated using [LaBSE](https://huggingface.co/sentence-transformers/LaBSE).
74
- The filtered datasets are then concatenated to form the final corpus.
 
 
 
75
 
76
  #### Who are the source language producers?
77
 
 
69
  The dataset is a combination of the following original datasets collected from [Opus](https://opus.nlpl.eu/):
70
  NLLB, MultiCCAligned, WikiMatrix, GNOME, KDE4, OpenSubtitles, GlobalVoices.
71
 
72
+ All data was filtered according to two specific criteria:
73
+ - Alignment: sentence level alignments were calculated using [LaBSE](https://huggingface.co/sentence-transformers/LaBSE) and sentence pairs with a score below 0.75 were discarded.
74
+
75
+ - Language identification: the probability of being the target language was calculated using [Lingua.py](https://github.com/pemistahl/lingua-py) and sentences with a language probability score below 0.5 were discarded.
76
+
77
+ The filtered datasets are then concatenated and deduplicated to form the final corpus.
78
 
79
  #### Who are the source language producers?
80