Community Audit: Amazigh languages
Here, we discuss about Amazigh languages (codes zgh, tzm, kab, shi ...).
@ayymen
Do you have any estimation of what the amount of these languages should be?
I want to share our non-clean data to see if we over-cleaned it with filters.
By the way, I very much use your reference for the correct labels for zgh and ary in Flores data.
Are zgh and tzm in Tifinagh very close?
@ayymen Do you have any estimation of what the amount of these languages should be?
Combined, they should have at least 10s or 100s of thousands of documents. kab, shi, and zgh have Wikipedias and many websites (government, news, etc...).
A useful heuristic is any Moroccan government website (.ma) page in Tifinagh is zgh, and any Algerian government website (.dz) page in Latin (but not in French) is kab-Latn.
I've noticed that there is a tzm-Tfng subset but not a tzm-Latn, and vice versa for shi. It actually should be the other way around, I know only one blog in tzm-Tfng, most of the content outside Wikipedia is probably in Latin and Arabic. shi on the other hand has a Wikipedia in Tifinagh and other websites.
I've also noticed that every page under taq-Tfng (and most under tzm-Tfng) is zgh-Tfng. This shouldn't happen for taq because zgh and taq don't use the exact same Tifinagh character set, so a LID model should easily tell them apart.
When I'll have more time I will play with GlotLID to see exactly what the model gets wrong.
Are zgh and tzm in Tifinagh very close?
Yes. zgh, shi, and tzm are very close (zgh being practically a superset of the other two).
our tzm_Tfng data comes from https://incubator.wikimedia.org/wiki/Wp/tzm
Is this a clean source for this language?our taq_Tfng major data comes from nllbseed which i think is not clean.
- I moved the tzm_Tfng into zgh_Tfng moslty
and parts from https://incubator.wikimedia.org/wiki/Wp/tzm to tzm_Tfng and parts from https://incubator.wikimedia.org/wiki/Wp/rif to rif_Tfng.
- our shi_Latn data for language identification does not come from wikipedia, but most of the captured content are from wikipedia so i see this as a good sign.
We do not have shi in Tfng. Is there any data for this you can point us to?
The biggest problem here is that besides kab-Latn, shi-Latn is the only Berber language in Latin script that we have. I found many examples of tzm-Latn and rif-Latn in the data, which I moved to their proper places. I will include them in the language ID training data so that this doesn't happen again.
in summary now we have these datasets:
kab-Latn: we moved obvious tzm-Latn and rif-Latn
shi-Latn: we moved abvious rif-Latn
rif-Latn: created from kab-Latn and shi-Latn
tzm-Latn: created from kab-Latn
tzm-Tfng: we moved obvious rif-Tfng and zgh-Tfng
zgh-Tfng: created from tzm-Tfng and taq-Tfng
rif-Tfng: created from tzm-Tfng
we removed taq-Tfng.
- our tzm_Tfng data comes from https://incubator.wikimedia.org/wiki/Wp/tzm
Is this a clean source for this language?
It looks good to me.
- our taq_Tfng major data comes from nllbseed which i think is not clean.
Probably of no real consequence since there's probably no data for it.
- I moved the tzm_Tfng into zgh_Tfng moslty
and parts from https://incubator.wikimedia.org/wiki/Wp/tzm to tzm_Tfng and parts from https://incubator.wikimedia.org/wiki/Wp/rif to rif_Tfng.
The incubator for rif is contaminated with shi and probably kab too.
- our shi_Latn data for language identification does not come from wikipedia, but most of the captured content are from wikipedia so i see this as a good sign.
We do not have shi in Tfng. Is there any data for this you can point us to?
Actually shi Wikipedia has both scripts, you can change the script from a drop-down menu. Here's the home page.
I can't see all the splits because of errors in HF's dataset viewer. Do you have an estimate of the percentage of documents written in Tifinagh that were filtered out?
By the way, Kabyle is mislabeled in UDHR. It was given the code tzm when it should have been kab. I found the error in your LID dataset as well.
Thanks done: https://huggingface.co/datasets/cis-lmu/udhr-lid/viewer/default/test?q=kab_Latn