Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

AI, Wikipedia, and uncorrected machine translations of vulnerable languages

technologyreview.com

Post reply on HN
Print view View original

41–50 of 84 posts

Re: AI, Wikipedia, and uncorrected machine translations of vulnerable languages

#41

Post by thaumasiotes »

don-bright wrote: post #37
aucisson_masque wrote: post #10

Earlier quoted context omitted.

I see that this comment get downvoted but I think we can agree on the facts that languages, just like species, die while other flourish. And that's fine. Survival of the fittest, right ? Not enough people speaking Greenlandic, too complicated even for it's own population who would rather speak danish ? The very reason I'm speaking English is because it was forced military during the 19th century by the UK and since t…

We already see the 'best' LLMs switch between different languages while they are 'thinking'. It seems to me that the more languages it can 'think' in, the better off it will be. Different human languages have different concepts of time, numbers, nature, place, intention, relationships, and so forth and so on.

> Survival of the fittest on a long time horizon means the more diversity the better the survival rate will be.

This is just a misapplication of the analogy. For a language, "fitness" refers to similarity to whatever language is spoken by people relevant to you. Diversity is the worst quality a language can exhibit, and is the quality that causes dying languages to die.

There is no such concept as an external force coming in that certain languages handle better, allowing them to temporarily outcompete other languages. Existing pools of diversity are not protective against this, because it can't happen.

Also unlike genetic diversity, linguistic diversity does not need to be maintained as a legacy of the past. It is constantly being generated in much larger quantities than are desired. If you managed to perform the opposite of the Tower of Babel miracle and replaced every currently-spoken language everywhere in the world with a perfect monoculture, within 1-2 generations you'd be back to having mutually unintelligible varieties in different regions.

Re: AI, Wikipedia, and uncorrected machine translations of vulnerable languages

#42

Post by orbital-decay »

That happens by default in low-resource languages, no bad translations needed. They don't have enough either written material to train an LLM, or labels for time periods and various dialects in a continuum. For example even the best multilanguage models will lump up all Berber languages into one unstable abomination nobody speaks, usually writing it in Neo-Tifinagh. Not much can be done about that, training a model in all varieties of these would require a huge specialized effort.

Re: AI, Wikipedia, and uncorrected machine translations of vulnerable languages

#43

Post by phantomathkg »

foxglacier wrote: post #2

If nobody's reading them and nobody's writing them, then perhaps it doesn't matter. We could let Wikipedia-Greenlandic persist as its own evolved language that forks from the original. > potentially pushing the most vulnerable languages on Earth toward the precipice as future generations begin to turn away from them. OK? We have lots of dead languages. It's fine. People use whatever languages are appropriate to them…

I thought the argument point is letting it die is OK, but letting wrongfully translated text, becoming the source of AI to chuck out wrongfully translated text is not OK.

Re: AI, Wikipedia, and uncorrected machine translations of vulnerable languages

#44

Post by bradley13 »

I've lived in a couple of countries where there is a "vulnerable" language. I understand the emotional attachment that the native speakers have to their language.

However, in the larger picture: languages evolve. New ones develop, old ones die. Do artificial attempts to "rescue" a language really make sense?

Re: AI, Wikipedia, and uncorrected machine translations of vulnerable languages

#45

Post by tdeck »

bradley13 wrote: post #44

I've lived in a couple of countries where there is a "vulnerable" language. I understand the emotional attachment that the native speakers have to their language. However, in the larger picture: languages evolve. New ones develop, old ones die. Do artificial attempts to "rescue" a language really make sense?

It makes no less sense than any other work done to protect or restore something created by human beings. This comment is no more insightful than saying "cathedrals burn down, do artificial attempts to 'restore them' make sense?"

Re: AI, Wikipedia, and uncorrected machine translations of vulnerable languages

#46

Post by tdeck »

aucisson_masque wrote: post #6

> Wehr, who now teaches Greenlandic in Denmark, speculates that perhaps only one or two Greenlanders had ever contributed. That's the core issue, it's not those who use AI translator or worst like Google translate. If there isn't any Greenlander to contribute to their Wikipedia, they don't deserve to have one and instead must rely on other languages. The difference between an empty Wikipedia and one filled with trans…

It is worse. Imagine if you were trying to learn English from this phrasebook, written by someone who didn't speak English:

https://www.exclassics.com/espoke/espkpdf.pdf

Wikipedia is prominent. Wikipedia articles in a language without much representation become prime examples of that language to those who read them.

Re: AI, Wikipedia, and uncorrected machine translations of vulnerable languages

#47

Post by debtta »

simonw wrote: post #4

I'm surprised this story didn't mention the scandal with Scots Wikipedia: https://www.theguardian.com/uk-news/2020/aug/26/shock-an-aw-... > an American teenager – who does not speak Scots, the language of Robert Burns – has been revealed as responsible for almost half of the entries on the Scots language version of Wikipedia It wasn't malicious either, it was someone who started editing Wikipedia at 12 and naively fa…

The background here is that Scots is not really a language. Try asking a Glasgow taxi driver who addresses you in 'Scots' whether he knows any English. Robert Burns wrote in English, with some of his spelling reflecting pronunciation in the Scottish English dialect.

The people who want it to be considered as a language for political reasons cannot be bothered to translate Wikipedia themselves. They read and edit English Wikipedia and understand it perfectly.

Re: AI, Wikipedia, and uncorrected machine translations of vulnerable languages

#48

Post by 1f60c »

orbital-decay wrote: post #42

That happens by default in low-resource languages, no bad translations needed. They don't have enough either written material to train an LLM, or labels for time periods and various dialects in a continuum. For example even the best multilanguage models will lump up all Berber languages into one unstable abomination nobody speaks, usually writing it in Neo-Tifinagh. Not much can be done about that, training a model i…

And it's a lot more profitable to improve sex mode than to hire a small army of native speakers to make it not suck at Greenlandic.

Re: AI, Wikipedia, and uncorrected machine translations of vulnerable languages

#49

Post by yorwba »

aucisson_masque wrote: post #6

> Wehr, who now teaches Greenlandic in Denmark, speculates that perhaps only one or two Greenlanders had ever contributed. That's the core issue, it's not those who use AI translator or worst like Google translate. If there isn't any Greenlander to contribute to their Wikipedia, they don't deserve to have one and instead must rely on other languages. The difference between an empty Wikipedia and one filled with trans…

"Only one or two" isn't zero. The problem isn't that a small community can only write a small Wikipedia, but that there's a global supply of fools who want to make every small Wikipedia bigger, even if they're completely unqualified to do so.

Wikipedia is built around the basic principle that if you just let everyone contribute, most contributions will be helpful and you can just revert the bad ones after the fact. This works for large communities that easily outnumber the global supply of fools, but below a certain size threshold, the sign flips and the average edit makes that version of Wikipedia worse rather than better.

So smaller communities probably need to flip the operating principle of Wikipedia on its head and limit new users to only creating drafts, on the assumption that most will be useless, and an admin can accept the good ones after the fact.

I'm not sure whether Wikipedia already has the software features necessary to operate it in such a closed-by-default manner.

Re: AI, Wikipedia, and uncorrected machine translations of vulnerable languages

#50

Post by mort96 »

bradley13 wrote: post #44

I've lived in a couple of countries where there is a "vulnerable" language. I understand the emotional attachment that the native speakers have to their language. However, in the larger picture: languages evolve. New ones develop, old ones die. Do artificial attempts to "rescue" a language really make sense?

Languages evolve, but it's probably bad when language evolution is driven by bad AI slop translations made by people who have no relation to the language.
Post reply on HN