I wonder of you pre train on Hebrew and Arabic if it will find the similarities between the RTL writing direction. So many similar words. I guess both came from Aramaic? If so, how about the trifecta of ancient languages with Aramaic then Hebrew the Arabic.
Do we need language specific LLMs? I can’t vouch for the data coverage or accuracy of Arabic in the leading models today, but I do know them to be highly cross-lingual capable.
Back when gpt 3.5 was released I was testing it for translating Tongan. A language that wasn't even on Google. It was doing okay. It lacks certain formalities and sort of contextual understandings in other languages like Tagalog or Spanish. But I noticed that if I put it into a character that is a native Filipino or Honduran. Or Tongan. It does do better. But gpt 5 really is leaps of evolution ahead compared to 3.5. It actually does really well now.
> Clear examples emerge when global language models address culturally sensitive issues, such as social relationships or political debates. They often adopt ambiguous positions that overlook the Arab cultural context, creating a gap between these digital tools and the values and lived experiences of Arab users.
Well I have bad news, my friend. English language models are also terrible at this.
This whole article seems to stem from the premise that it's important for LLMs to engage cultural issues competently. But... should they even?
7 comments
[ 3.5 ms ] story [ 29.0 ms ] threadWell I have bad news, my friend. English language models are also terrible at this.
This whole article seems to stem from the premise that it's important for LLMs to engage cultural issues competently. But... should they even?