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

51–60 of 84 posts

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

#51

Post by orbital-decay »

1f60c wrote: post #48
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.

What makes Greenlandic special among ~7000 languages in the world? Most of them are low-resource as well. To train a model in all of them you also need a ton of specialized linguists and ML people, neither of which grow on trees. And it's only one thing generalist models are supposed to master, out of many. The scale is impossible, this needs to be done by models themselves when (if) they get smart enough.

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

#52

Post by carlosjobim »

AlienRobot wrote: post #15
ratg13 wrote: post #3

It's ironic that the "solution" to the problem is being driven by yet another person that isn't native to Greenland. While they may be a Greenlandic teacher, it's almost assured that they are teaching western Greenlandic, which is similar to Canadian Inuktitut. People in the East of Greenland speak a language that has similarities, but is different enough in vocabulary and sounds that it's often considered a separate…

As someone who isn't a native English speaker, I believe most people who use the Internet would benefit from simply learning English rather than having an unchecked AI translate things to them. Reddit for example has joined millions of terrible Wordpress websites in auto-translating everything for SEO purposes and Google seems to be fine with this for some reason. It's ironic that it has reached the point that if you…

You miss an advantage. If everything is inter-translated, then you can do your search in the language you know and find the answer written in a language you didn't know.

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

#54

Post by Pooge »

Hendrikto wrote: post #53

Wow, fuck that site. I had to dismiss 2 cookie banners and 3 popups before I could even read anything, then the second I scrolled one pixel another one popped up.

Sounds like you don't use the modern-day condom: uBlock Origin.

I live in Europe (famous place for becoming professional banner clicker) and I didn't get one single distraction.

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

#56

Post by zozbot234 »

debtta wrote: post #47
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 Engl…

Scots is somewhat partially intelligible in written form to English speakers, but that does not make it the same language as English. You might as well say that Spanish and Portuguese are the same language.

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

#57

Post by Hendrikto »

Pooge wrote: post #54
Hendrikto wrote: post #53

Wow, fuck that site. I had to dismiss 2 cookie banners and 3 popups before I could even read anything, then the second I scrolled one pixel another one popped up.

Sounds like you don't use the modern-day condom: uBlock Origin. I live in Europe (famous place for becoming professional banner clicker) and I didn't get one single distraction.

I am on mobile, I should have said that.

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

#58

Post by r2vcap »

While it makes sense that LLMs and machine translation systems primarily rely on English Wikipedia as a data source, depending on smaller-language Wikipedias for training is far less ideal. English Wikipedia is generally well-regulated by its community, but many other language editions are not — so treating all of Wikipedia as an authoritative source is misguided.

For instance, my mother tongue’s Wikipedia (Korean Wikipedia) suffers from serious governance issues. The community often rejects outside contributors, and many experienced editors have already moved to alternative platforms. As a result, I sometimes get mixed, low-quality responses in my native language when using LLMs.

Ultimately, we need high-quality open data. Yet most Korean-language content is locked behind walled gardens run by chaebols like Naver and Kakao — and now they’re lobbying the government to fund their own “sovereign AI” projects. It’s a lose-lose situation.

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

#59

Post by mikemcquaid »

debtta wrote: post #47
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 Engl…

Sort of?

The Glaswegian taxi driver may not consider themself to be speaking a different language but, if speaking to another local and leaving aside pronunciation, they’d use words, phrases and even grammar that’s incomprehensible to someone with no experience with Scots.

I’m a “posh Scot”, raised middle class in Edinburgh so my accent is minimal and thickens up or softens depending on who I’m speaking to. Even for me, there’s a lot of words, phrases and ways of speaking I’ve had to adjust to be consistently understood by American coworkers when over the last 10+ years.

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

#60

Post by user205738 »

Hendrikto wrote: post #57
Pooge wrote: post #54

Earlier quoted context omitted.

Sounds like you don't use the modern-day condom: uBlock Origin. I live in Europe (famous place for becoming professional banner clicker) and I didn't get one single distraction.

I am on mobile, I should have said that.

ublock origin works on mobile in Firefox, Waterfox, etc.

There is Brave with its blocker, there is AdGuard, which blocks ads on websites and applications, regardless of the browser.

Post reply on HN