I'm a simultaneous interpreter, and it's a bloodbath out there. Partially because anyone who needs a translator or an interpreter by default is unable to verify the accuracy of the translation/interpretation - they can only tell if it's smooth, believable, and such. And, AIs are great at being believable even when they make shit up...
olympicyes@lemmy.world
on 27 Aug 01:35
nextcollapse
I tested a few local models to see how complete and recent their training data is. I want to use it to see if company A at xyz address is the same as Company B at xyz.1 address. I asked them for recent events and found a lot of gaps. So I asked for the roster of the 1992 dream team. Very mixed results. Open AIs model got 11/12 players correct but absolutely insisted that Christian Laettner was not the 12th player. I went back and forth with it to see if I could get it to accept my knowledge as is. It wouldn’t. I’m terrified about what happens when these AI bots have the ability to update Wikipedia in order to make the facts match their incomplete training data.
You don't have to be very good at a language to know when a translation is horrible. I'm not very good at spanish and I can do better than machine translations.
tiredofsametab@fedia.io
on 26 Aug 23:08
nextcollapse
My company thankfully still employs simultaneous interpreters for meetings and has one translator on staff. I think, at least in part, because of how bad translation tools can be from EN <> JA.
OmegaSunkey@ani.social
on 27 Aug 00:17
nextcollapse
I believe this will pass. Sooner or later, the AI companies will have to stop losing money and adjust their pricing. And then it’ll turn out that using AI for everything gets you worse results than humans, at the same cost. And that will be that. I hope I can hang on until then.
such hope, i wish i had
AdamEatsAss@lemmy.world
on 27 Aug 01:34
nextcollapse
Or humanity will run out of power and the servers required to run AI will power down.
scintilla@crust.piefed.social
on 27 Aug 01:36
nextcollapse
I think the AI bubble is really close to poping. If they can't make a profit on it in the next 2 years (maybe sooner with how badly the economy is about to start doing) VC money will dry up and then they will have to up prices dramatically.
There's a lot of factors I could get into but it seems like they are going to have to start pricing AI use at the very least cost soon.
It really sounds like a "Let's go to the Winchester, have a nice cold pint, and wait for this all to blow over" reaction.
Things will be different. Maybe AI won't replace humans entirely in this role, but it's still going to be a major tool in the toolbox.
elucubra@piefed.social
on 27 Aug 08:25
nextcollapse
What most people managing translations don't get is that they are essentially using the tools that translators use, but skipping the value adding step.
I've been doing translation as a side gig for years. Lately I've been doing some translations for an NGO that deals with addiction management, of which I'm part.
The materials have a lot of nuances, and need the translator to understand them, to properly convey the concepts.
The usual process for translation is to feed the original to a machine language translation software, and then work with both versions side by side, in a translation management software, tools that make editing and proofing faster and easier by a human, to achieve the best result.
Last time, someone in the organization, mono lingual, decided to do a handbook translation with ChatGPT, or something like that.
They then gave the result to a colleague and me.
The resulting translation was exactly what we expected.
A problem was that some bilingual people were shown the results, and reported that the results were amazing, without realizing that they were commenting on the wow factor, not on the accuracy of the result, especially because they had not done a critical side by side comparison.
My colleague and I did the editing work, were paid less, but the end result was the usual translation quality.
The commissioning person at the org boasted that AI translation was great, obviating our work, to get their brownie points.
TLDR: translation has used machine translation as a first step for a long time, with results edited and polished by humans. Ignorant decision makers are skipping that crucial step, getting sub-par results, oblivious to the fact.
zipzoopaboop@lemmynsfw.com
on 27 Aug 23:08
collapse
I would consider what you do localization. Translations are 1 aspect of that role
Not really our case. We do English->Spanish, where we try to achieve the most neutral Spanish, as there are many local variations. Think truck/lorry, for example.
It's more translating expressions or phrases that don't convey the same concept. For example, "by the way" could be translated to "por el camino" which doesn't usually have the same usage.
zipzoopaboop@lemmynsfw.com
on 29 Aug 17:52
collapse
Exactly, not just literal translation but an aspect of art in conveying the meaning and intent of the writer
I’ve always thought a BASIC international language would be great. I mean, there’s already international sign language, and Arabic numerals are pretty universal. Doesn’t need to be poetic, or intense. Just “Me want this, I need this” type of structure. Maybe a modified version of Latin, with all gender neutral variants.
Yeah it’s totally worth destroying most cultures on the planet to improve water usage just long enough for the machine to find another way to waste it. Terrible take.
I know someone who was a translator between two (less widely spoken) languages, and some specifics I recall from our conversations about work:
Sometimes the translations use many technical terms, and getting those wrong (trusting LLMs) is not an option. (This was for some patents IIRC)
Some terms simply do not exist in another language, and it could be up to the translator to invent a term to define and carry the information across. (This was for some government digital service, and the term was similar to “digital queue”)
Tone and nuances are very difficult to translate. Phrasing can have implications and connotations. (Simplest example: “i am afraid” does not imply fear, it’s an established politeness phrase) Neutral in one language could be viewed as hostile in another, too. (And with politicians being petty, could have consequences)
None of those would be addressed with LLMs. Small training set for language (and language being similar to a few others) is an issue. Anything technical or non-existing would be prone to hallucinations. And tone is difficult enough to convey through text to begin with, let alone with LLM translation.
ExLisper@lemmy.curiana.net
on 27 Aug 15:27
nextcollapse
I wonder what % of all translations are things like patents, legal paper and movies and what are simple localizations. Even in the more complex cases you can pass the entire text through AI first and then just proof read it and correct the errors.
vacuumflower@lemmy.sdf.org
on 28 Aug 10:47
collapse
That proofreading is as hard as with code. Defeats the purpose.
Unfortunately it doesn’t have to be better than the worker, we all know this sucks at most of the things it’s being touted as great at.
It just has to convince management who make decisions that it’ll save money (or that they can spin it that way) for the next quarter. That alone is enough to destroy people’s lives.
OmegaLemmy@discuss.online
on 29 Aug 08:13
collapse
LLM gets 95% of the translation done, but the 5% is likely every important and it takes longer to confirm it’s correct than to do it from scratch anyway
How good is LLM training data for a language spoken by less than 10 million people? Keep in mind that most of those people are probably multilingual (i.e. categorizing which language is which by person is harder), and language itself is similar to its neighbors. And then, again, terms.
OmegaLemmy@discuss.online
on 30 Aug 15:21
collapse
I can not say, and wouldn’t trust it unless a translator confirmed its validity
I’ve had to translate a whole bunch of letters from English to Finnish for my grandparents, and doing it using a translator saves a ton of time as I don’t have to actually produce the text, I can just read both sides afterwards and as long as every sentence matches in meaning, I can move to the next one.
But I wouldn’t trust it to actually be correct for the entire thing, because it never is, and if someone who doesn’t understand one of the languages would do it they would never spot the mistakes either.
vacuumflower@lemmy.sdf.org
on 28 Aug 10:35
collapse
I’d expect any translation requiring zero mistakes and translator’s official responsibility wouldn’t be hurt by this.
threaded - newest
It‘s terrible and sad. Even more so because AI still gets things wrong all the time.
I'm a simultaneous interpreter, and it's a bloodbath out there. Partially because anyone who needs a translator or an interpreter by default is unable to verify the accuracy of the translation/interpretation - they can only tell if it's smooth, believable, and such. And, AIs are great at being believable even when they make shit up...
I tested a few local models to see how complete and recent their training data is. I want to use it to see if company A at xyz address is the same as Company B at xyz.1 address. I asked them for recent events and found a lot of gaps. So I asked for the roster of the 1992 dream team. Very mixed results. Open AIs model got 11/12 players correct but absolutely insisted that Christian Laettner was not the 12th player. I went back and forth with it to see if I could get it to accept my knowledge as is. It wouldn’t. I’m terrified about what happens when these AI bots have the ability to update Wikipedia in order to make the facts match their incomplete training data.
You don't have to be very good at a language to know when a translation is horrible. I'm not very good at spanish and I can do better than machine translations.
My company thankfully still employs simultaneous interpreters for meetings and has one translator on staff. I think, at least in part, because of how bad translation tools can be from EN <> JA.
such hope, i wish i had
Or humanity will run out of power and the servers required to run AI will power down.
I think the AI bubble is really close to poping. If they can't make a profit on it in the next 2 years (maybe sooner with how badly the economy is about to start doing) VC money will dry up and then they will have to up prices dramatically.
There's a lot of factors I could get into but it seems like they are going to have to start pricing AI use at the very least cost soon.
It really sounds like a "Let's go to the Winchester, have a nice cold pint, and wait for this all to blow over" reaction.
Things will be different. Maybe AI won't replace humans entirely in this role, but it's still going to be a major tool in the toolbox.
What most people managing translations don't get is that they are essentially using the tools that translators use, but skipping the value adding step.
I've been doing translation as a side gig for years. Lately I've been doing some translations for an NGO that deals with addiction management, of which I'm part.
The materials have a lot of nuances, and need the translator to understand them, to properly convey the concepts.
The usual process for translation is to feed the original to a machine language translation software, and then work with both versions side by side, in a translation management software, tools that make editing and proofing faster and easier by a human, to achieve the best result.
Last time, someone in the organization, mono lingual, decided to do a handbook translation with ChatGPT, or something like that.
They then gave the result to a colleague and me.
The resulting translation was exactly what we expected.
A problem was that some bilingual people were shown the results, and reported that the results were amazing, without realizing that they were commenting on the wow factor, not on the accuracy of the result, especially because they had not done a critical side by side comparison.
My colleague and I did the editing work, were paid less, but the end result was the usual translation quality.
The commissioning person at the org boasted that AI translation was great, obviating our work, to get their brownie points.
TLDR: translation has used machine translation as a first step for a long time, with results edited and polished by humans. Ignorant decision makers are skipping that crucial step, getting sub-par results, oblivious to the fact.
I would consider what you do localization. Translations are 1 aspect of that role
Not really our case. We do English->Spanish, where we try to achieve the most neutral Spanish, as there are many local variations. Think truck/lorry, for example.
It's more translating expressions or phrases that don't convey the same concept. For example, "by the way" could be translated to "por el camino" which doesn't usually have the same usage.
Exactly, not just literal translation but an aspect of art in conveying the meaning and intent of the writer
The world would be soo easy if there was 1 lang. No more translations no more burning money water electricity
Also one clothes, one house and one grave. I think I’ve seen this before.
Where?
In Auschwitz
I’ve always thought a BASIC international language would be great. I mean, there’s already international sign language, and Arabic numerals are pretty universal. Doesn’t need to be poetic, or intense. Just “Me want this, I need this” type of structure. Maybe a modified version of Latin, with all gender neutral variants.
That’s English to a big chunk of the world already.
Esperanto exists.
Yeah it’s totally worth destroying most cultures on the planet to improve water usage just long enough for the machine to find another way to waste it. Terrible take.
Not to mention the blow to the nationalism.
I know someone who was a translator between two (less widely spoken) languages, and some specifics I recall from our conversations about work:
None of those would be addressed with LLMs. Small training set for language (and language being similar to a few others) is an issue. Anything technical or non-existing would be prone to hallucinations. And tone is difficult enough to convey through text to begin with, let alone with LLM translation.
I wonder what % of all translations are things like patents, legal paper and movies and what are simple localizations. Even in the more complex cases you can pass the entire text through AI first and then just proof read it and correct the errors.
That proofreading is as hard as with code. Defeats the purpose.
Unfortunately it doesn’t have to be better than the worker, we all know this sucks at most of the things it’s being touted as great at.
It just has to convince management who make decisions that it’ll save money (or that they can spin it that way) for the next quarter. That alone is enough to destroy people’s lives.
LLM gets 95% of the translation done, but the 5% is likely every important and it takes longer to confirm it’s correct than to do it from scratch anyway
How good is LLM training data for a language spoken by less than 10 million people? Keep in mind that most of those people are probably multilingual (i.e. categorizing which language is which by person is harder), and language itself is similar to its neighbors. And then, again, terms.
I can not say, and wouldn’t trust it unless a translator confirmed its validity
I’ve had to translate a whole bunch of letters from English to Finnish for my grandparents, and doing it using a translator saves a ton of time as I don’t have to actually produce the text, I can just read both sides afterwards and as long as every sentence matches in meaning, I can move to the next one.
But I wouldn’t trust it to actually be correct for the entire thing, because it never is, and if someone who doesn’t understand one of the languages would do it they would never spot the mistakes either.
I’d expect any translation requiring zero mistakes and translator’s official responsibility wouldn’t be hurt by this.