Any updates coming?
Is there any plan on updating the dataset with new translations that may have been added in the last 2 years? I was planning on doing this dataset myself but found it already exists. I may try to do a new export from weblate if this dataset does not get any update.
It's been very helpful for training models for https://mozilla.github.io/translations/
Hello @ZJaume , I'm glad you found the dataset useful ๐ if I knew it was being used at Mozilla I would have kept it updated at least on a quarterly basis. Looking now at their API, it seems the dataset has indeed grown by roughly 52%. Hopefully there are also new languages there. I have a couple of questions if you don't mind:
- The first time I uploaded the data I tried to clean up the codes but quickly gave up, there's just so much stuff here - pirate speak, "uwu" speak, emoji language, multiple subsets for the same language, etc... so I'd like to know your thoughts in case you've looked at the different subsets or compared them. For example, are the various suffixes attached to lang codes informative? for instance in the
en-dedirection there's@ formal,@ informal,_1901,_rudeand many others. Your feedback can help me simplify the codes and improve documentation to make the dataset easier to use. - Are you interested in using other multilingual localization datasets such as this one? there are other data sources as well.
I will let you know once I upload a new version.
Great news then!
Regarding the questions:
- The language codes seem pretty messy, but taking into account that for the majority of the languages, 90% of the data is in the en-xx code, there's not much need in cleaning that up. In the future, we may train more low-resource languages that might raise issues about the language code mess, so I guess I could try to report them to you in the future. But anyway, I made a Huggingface dataset importer for the pipeline that let's me add several subsets of a dataset if needed, so I can, let's say, add
en-ptand all other subsets with country BCP code suffixes in the same pipeline configuration without too much trouble. - Yes, we are interested in more localization datasets. But small datasets like the one you are pointing out are much very important for now (if they have only a few thousand pairs per language, for example 1.3k sentences for Hungarian won't make much difference), until we start training very low resource languages, which may take some time until we start.
- I was also wondering if you used any of the metadata available in Weblate. The exported TMX has a field that tells if the translation has been finished and that could work as double quality check. Although I'm not sure if can be used if decreases each language pair data by 20% or more. I guess it will depend on the quality of discarded sentence pairs, if a lot of bad translations get discarded, it's good. Maybe it could be added as a dataset column.
Indeed, we are interested in any kind of parallel data, not just localization data.