Google Translation AI Botches Legal Terms 'Enjoin,' 'Garnish' (reuters.com) 84
Translation tools from Google and other companies could be contributing to significant misunderstanding of legal terms with conflicting meanings such as "enjoin," according to research due to be presented at an academic workshop on Monday. From a report: Google's translation software turns an English sentence about a court enjoining violence, or banning it, into one in the Indian language of Kannada that implies the court ordered violence, according to the new study. "Enjoin" can refer to either promoting or restraining an action. Mistranslations also arise with other contronyms, or words with contradictory meanings depending on context, including "all over," "eventual" and "garnish," the paper said.
Google said machine translation is "is still just a complement to specialized professional translation" and that it is "continually researching improvements, from better handling ambiguous language, to mitigating bias, to making large quality gains for under-resourced languages." The study's findings add to scrutiny of automated translations generated by artificial intelligence software. Researchers previously have found programs that learn translations by studying non-diverse text perpetuate historical gender biases, such as associating "doctor" with "he." The new paper raises concerns about a popular method companies use to broaden the vocabulary of their translation software. They translate foreign text into English and then back into the foreign language, aiming to teach the software to associate similar ways of saying the same phrase.
Google said machine translation is "is still just a complement to specialized professional translation" and that it is "continually researching improvements, from better handling ambiguous language, to mitigating bias, to making large quality gains for under-resourced languages." The study's findings add to scrutiny of automated translations generated by artificial intelligence software. Researchers previously have found programs that learn translations by studying non-diverse text perpetuate historical gender biases, such as associating "doctor" with "he." The new paper raises concerns about a popular method companies use to broaden the vocabulary of their translation software. They translate foreign text into English and then back into the foreign language, aiming to teach the software to associate similar ways of saying the same phrase.
Sanctioned violence (Score:3)
Inflammable materials
An oversight
It ain't the google that's the problem, it's the messiness in colloquial speech that's the problem. This is why specialized domains (law, medicine, physics, etc) have specific meanings for certain words that may or may not match the common meanings of those same words. Usually the meaning is related by metaphor, but it's not a good idea to make an assumption without consulting a specialized dictionary.
Re: (Score:3)
Kinda reminds me of the 1983 Article [nytimes.com] attempt to develop translation software to convert between English and Russian.
The developers got a little cute with the testing, entering "The spirit is willing, but the flesh is weak" as the English starting phrase. This was translated to Russian and the output translated back to English.
The result?
"The vodka is strong, but the meat is rotten."
What's interesting about this is that you can't really rely on literal translation techniques for colloqui
Re: (Score:2)
Which is why I like the pre-fabricated words and sentences translated across five languages in Final Fantasy XIV. Even if people get "creative" with it, there's no computer trying to interpret and translate things by itself.
Re: (Score:3)
Re:Sanctioned violence (Score:4, Informative)
Travelers use phrasebooks for the same reason.
My hovercraft is full of eels.
Re: (Score:2)
Re: (Score:2)
Menus often have spelling or grammar mistakes, especially when the menu wasn't created by natives of the language - which is common for small shops (employing cheap labor) or restaurants specialising in a foreign cuisine (often employ staff from the country where the cuisine originates).
Couple this with translation and you can have some craziness...
Recently i've seen:
Mined chicken
Lamp curry
Soft shell crap
Re: (Score:2)
lol
Re: (Score:2)
This is something different. Google's new AI-based translation gets antonyms mixed up fairly often. I don't know why. I've seen it mix up "Woman" and "man."
Here's another example [upenn.edu], that people thought was China influencing Google, but it's not. It's a normal type of error in the system. Antonyms often show up in the exact same contextual sentences (with no other words changed), so context based translation gets confused about which word to show.
Re: (Score:2)
Humans occasionally do this too! E.g., people who think "nonplussed" means "unfazed".
Re: Sanctioned violence (Score:2)
When humans do it, language evolves. When machines do it, it's clearly a mistake.
This thinking is racist against cylons.
Re: (Score:2)
The translator should be able to detect that it is translating a legal document, or at least a legal discussion, and then make the appropriate domain-specific translation.
If the surrounding text talks about a court order, delinquent child-support, and paychecks, then "garnish" is unlikely to refer to parsley.
Re: (Score:2)
My favourite example:
An Italian text with a list of countries, including Macedonia, a country to the north of Greece, which has since been renamed as North Macedonia. The region of Greece it borders is also called Macedonia.
Google translated it to "Fruit Salad". If you go into a restaurant in Italy and ask for "macedonia", they will bring a fruit salad to the table. But Google wasn't smart enough to realise that if it is included in a list along with things like "Serbia", "Bosnia", and "Grecia", it is proba
Re: (Score:2)
A long, long time ago I remember reading a sci-fi book about an astronaut who was thrown out of NASA and charged with manslaughter after he actively took steps to move a vehicle (Shuttle Orbiter?) in order to avoid debris. In the actual text of the book, there were various astronauts screaming instructions to each other in a heightened state of panic. The guy at the controls wanted conf
Re: (Score:2)
The problem is that Google Translate, and all other translation tools I have tried, don't understand context.
You get similar issues with Japanese quite often. Sentences don't include information about who is speaking or acting, it's just assumed that humans will know from context. Often headlines get translated as if someone was describing something they did, e.g. "I tasted the new Muscat flavour ice-cream" instead of "New Muscat flavour ice-cream tasting".
Gender is often not stated either and Google doesn'
Re: (Score:2)
The real problem in context is that often the context is implied. That is something that current translation software often fails at as the context might be defined 10 sentences away by a side not.
Thus I consider the proper limit of General artificial intelligence to be "When can an AI understand and follow the context and context switches of a free flowing conversation."
Re: (Score:2)
missile e... "side note".
Re: Sanctioned violence (Score:3, Insightful)
Re: Sanctioned violence (Score:1)
There are (for example) medical datasets. And google translate seems to do a good enough job translating medical lab reports from Mandarin into English for a doctor here to understand what's intended. But that is the key: specialist understanding of context on the receiving end and commonality of concepts. Medicine is medicine everywhere, but different versions of English Common Law are different enough as it is, nevermind the occasional lack of 1 to 1 correspondence between American legal concepts and Camb
What does "Understanding" mean? (Score:2)
AI translaters do understand at some level. Those patterns are a form of understanding. And probably deeper than some human translators of technical texts.
And generally, humans tend to focus on words rather than meanings. "Weapons of Mass Destruction" comes to mind -- what did it actually mean?
When people criticize AIs for not being intelligent, they forget that humans are not that intelligent either.
Re: (Score:1)
Weapons of Mass Destruction
Simple: they destroy mass (and create energy).
Re: (Score:2)
Re: (Score:2)
Translation tools do understand grammar and sentence structure.
Re: (Score:2)
Gender is often not stated either and Google doesn't seem to recognize the gender of names, so you get things like "Naomi Osaka won his 3rd tournament".
And if google tries to recognise the gender of names, they will be attacked by SJWs for getting the gender wrong.
Re: (Score:2)
Trying seems better than just assuming everyone is male, which is what it does now.
Re: Sanctioned violence (Score:2)
Didn't you get the memo? Trying would run the risk of misgendering. See back in the day, people would pick names from gendered lists. Many languages have masculine and feminine versions of many commonly used names in common use among both genders. This used to be the preferred compact method of indicating gender. So you could use machines to infer the correct pronouns (and in some languages, verb conjugations and adjective forms). But now, any text that doesn't blast out the preferred pronouns of the person
Re: (Score:2)
Misgendering is what it does now. At least they could get points for trying.
The worst bit is when it assumed gender based on job title.
Re: (Score:2)
Another problem there is, in English-speaking countries, Andrea is a name that is usually given to girls. In Italy, it is usually given to boys. But, in Italy, the usual rule is that names ending in o are given to boys and names ending in a are given to girls. But Luca and Andrea are two exceptions to that rule.
Re: (Score:2)
Names like Vivian can be used for girls or boys in English. Maybe the solution for English is to use gender neutral language.
In Japanese there is a larger difference between the way men and women speak than in most European languages. It sounds quite odd when you have a man speaking in a feminine way. Of course that opens up a whole other debate about gender conformity in Japanese society...
Re: (Score:2)
Inflammable
Another word that should be taken out back and shot.
Re: Sanctioned violence (Score:3)
Careful. 50/50 it blows up in your face.
Inflammable means flammable? What a country! (Score:3)
Judge Harm: Further, I hereby garnish Homer's wages until Bart is fully repaid.
Homer: Mmm.. garnish.
[ Two more 'm's and this would be ASCII art! ]
Re: (Score:2)
The word is not at fault. English simply has two different in- prefixes. The original one means exactly "in", the other, imported from Latin, means "not". Why importing it when there exists a perfectly fine un- prefix, I do not know.
Inflammable is already banned (Score:2)
You will not see it on signs in Oz for some time. "This petrol says that it is inflammable, so it is OK to store it inside the boat"...
Re: (Score:2)
Inflammable
Another word that should be taken out back and shot.
It has already been shot. The word "inflammable" is banned by the military, commercial aviation, and many regulatory agencies.
It is a stupid, dangerous word, and there is no reason to use it. "Flammable" is shorter and clearer.
Re: Sanctioned violence (Score:2)
That comment might inflame linguistic strife.
Re: (Score:1)
so google should flag words with double+ meanings
it IS google that's the problem because "landing on the moon isn't difficult, it's just the the hard vacuum of space and that the moon is a rock with jaggy edges"... so what, that's what landing on the moon means.
you want to understand and/or translate languages, you have to shoulder the entire task, either you are Turing or turnt.
All that to say (Score:2)
Inflammable materials
An oversight
It ain't the google that's the problem, it's the messiness in colloquial speech that's the problem. This is why specialized domains (law, medicine, physics, etc) have specific meanings for certain words that may or may not match the common meanings of those same words. Usually the meaning is related by metaphor, but it's not a good idea to make an assumption without consulting a specialized dictionary.
Context Matters. Congradulations.
Our language is so bad... (Score:1)
Unbridled Hate (Score:5, Insightful)
Yes yes yes.
I hate "enjoin" almost as much as I hate the word "sanction".
A single word with two completely opposite meanings. HATE IT.
Also, "cleave", "overlook", and "screen". Stupid words that exist to confuse and obscure clarity.
"Raze" and "raise" come close but at least they're spelled differently.
Re: (Score:2)
It’s almost as if it is deliberate, at least in the legal arena. I can think of a few engineering terms that might be a 90-degree divergence from what someone might think based on language analysis, but at least it isn’t a complete 180!
Re: (Score:1)
Re: (Score:2)
And then there are "controversial" words such as niggardly. https://en.wikipedia.org/wiki/... [wikipedia.org]
It's only "controversial" to those who aren't educated. Sadly, there are a lot of those people.
Re: (Score:1)
You are talking about a field of endeavor that has officially decreed that "shall" means "may". They manufacture ambiguity to keep other lawyers busily employed. You could start out with entirely unambiguous language, and within a few court cases, you would have cases like "shall means 'may'".
Re: (Score:2)
You are talking about a field of endeavor that has officially decreed that "shall" means "may".
Not in my field. In my work, "shall" is a directive that doesn't mean "may", it means something that absolutely will be done, period. No ifs, ands, or buts, no conditionals, no wiggle room.
You may disagree, as I suspect you shall.
Re: (Score:2)
I'm an engineer. I had an argument with one of the old engineers here, who was relying on sources like this one [faa.gov] about "shall" versus "must":
The old guy eventually relented when everyone else on the team argu
Re: (Score:2)
Can you cite a source. Because it does not pass the smell test
Re: (Score:2)
Beyond the FAA page I cited before you posted that comment:
https://www.abajournal.com/mag... [abajournal.com]
https://www.plainlanguage.gov/... [plainlanguage.gov]
and Gutierrez de Martinez v. Lamagno, 515 U.S. 417 (1995).
Re: (Score:2)
Don't forget literally, which can literally mean figuratively.
Re: (Score:2)
Don't forget literally, which can literally mean figuratively.
Not for anyone who graduated from 5th grade.
"Literally" means "actually". It does not mean "sort of" or "kind of" or "not really".
You may disagree but a court of law will set you straight. Testifying that someone "literally knocked you down the stairs" when that didn't actually happen will get you a perjury charge.
Re: (Score:2)
Educate yourself:
https://www.oed.com/viewdictio... [oed.com]
https://www.merriam-webster.co... [merriam-webster.com]
You might want to rethink your snarky reply now. As OED points out, this usage has existed since at least 1769, so you are only about 250 year behind the times.
Re: (Score:2)
No, words have meaning. Anyone who cares about language and uses it properly will tell you you're wrong.
Again, testify under oath in court that someone "literally did X" when it didn't actually happen, and you'll face a perjury charge.
Go ahead, try it and then tell the judge that he's wrong.
But hey, you're literally a fucking asshole who rapes children, and in my colloquial usage that means you're just a great guy. Seriously.
Don't complain, I meant it literally. What's the problem?
Re: (Score:2)
Lol. Ok, you got really angry and swore at me so I guess you win automatically. From now on instead of going to the Oxford English Dictionary (formerly the most authoritative source on the history of the English language) to lookup the meaning of words, well just all go to JustAnotherOldGuy (the new authoritative source on the English language).
I'll get off your lawn now so you can resume yelling at clouds.
Re: (Score:2)
Lol. Ok, you got really angry and swore at me so I guess you win automatically.
Oh noes, did I offend your delicate sensibilities? You poor baby, do you need me to point you to the nearest fainting couch? Goodness gracious, I'm such a beast!
Why are you upset? I told you I meant it literally, so why are you offended? According to you it means the opposite, right?
Anyway, fuck you and your faux outrage.
I'll get off your lawn now so you can resume yelling at clouds.
You wouldn't be allowed on my lawn in the first place. And I mean that literally.
Re: (Score:2)
Lol. My argument is that nearly every dictionary defined the word that way. Your argument is, bahhh, what does a dictionary know about what words mean.
But mostly I'm just worried you forgot to take your blood pressure meds again, grandpa.
Re: (Score:2)
Don't worry, honey, I'll be fine. You, on the other hand, have ailments that can't be corrected by medication.
It's hilarious that you're literally losing your mind over this, lol.
Re: (Score:2)
And you're at least talking about words that are common to variants of English. When you look between variants of english you get different interpretations as well: "to table" means the exact opposite in US English as British English where the equivalent term would actually be "to shelve".
Or then you get into undefined prefixes such as bi- which as a prefix means both twice and every two times. We should revisit this bimonthly. Is that 6 times a year or 24 times?
Re: (Score:2)
Or then you get into undefined prefixes such as bi- which as a prefix means both twice and every two times. We should revisit this bimonthly. Is that 6 times a year or 24 times?
That's why I always correct stuff like that to "twice a week" or "every two weeks", after clarifying it with the original author. If I had a dollar for every time I've done that...
The same with the word "once", which people misuse with annoying regularity.
"Once the software is installed...." NO. No no no no no no. It's "after the software has been installed..."
You may laugh at this kind of nitpicking, but the fact is that these kinds of mistakes end up getting written in blood (just like safety regulations
Story of the talking cow. (Score:3)
When he narrated that incredible talking cow incident to the tow truck operator the tow truck guy started bitching about how stupid the cow was, to think diesel engines have carburetors.
The story is not what Google translate got wrong, but how much it got right. I have tested it in another Indian language, Tamil. I would say what it does is incredible. Not 100% accurate and very funny "original" and innovative interpretations for some phrases and sentences. But overall impressive, almost as impressive as a cow talking about carburetors of diesel engines.
Re: (Score:2)
The problem is that a straight dictionary translation would be better than the probabilistic/AI approach that Google translate currently takes.
Re: (Score:2)
Straight dictionary translations do not work at all. That was tried in the early days.
Re: (Score:2)
The problem is that a straight dictionary translation would be better than the probabilistic/AI approach that Google translate currently takes.
That is completely false since there are many words in both languages which take on various definitions depending on context.
You literally do not at all solve the problem being discussed while making the actual translation significantly worse as a result. Just occasionally I wish you'd think through the things you post on here before hitting submit.
It used to be that if you wanted a cheap (Score:1)
...shortcut and dump the side-effects on your successor after you got your Big Raise for "saving money", you outsourced it overseas*. Now you outsource to bots.
* It's not that overseas workers are incompetent, it's more often that they don't understand your business and don't want to make waves, so they quietly force round pegs into square holes.
Love my bot! (Score:1)
My garnished wages were delicious!
Re: (Score:1)
yeh... (Score:2)
Automatic translation, even with fancy pants machine learning system still not as good as humans at understanding meaning... News at 11.
deepl (Score:2)
I recently tried https://www.deepl.com/translat... [deepl.com] . I ran a few texts through it, including some news articles, a couple of science popularization fragments and even poetry. It was significantly better than Google's translation, and captured far more of the context. It wasn't perfect (especially the poem), but it was pretty close and sounded almost natural. By contrast, the same poem translated by Google was mostly gibberish, and the parts that weren't sounded really creepy.
Pretty Good, Not Perfect (Score:5, Interesting)
I am a lawyer for a company that does business in a Spanish speaking country. I'm often reading new laws or reading official government correspondence. I do have a pretty good working knowledge of Spanish such that I can read the original, but much more slowly and with a lot of stops to look up words that are not in common usage. Google translate speeds up the process quite a bit and is a really helpful tool. I can have it do the whole document, or just use it as a quick reference dictionary (especially useful for phrases).
That said, I would never use it as the basis of a legal opinion if there is any question about connotation or specific meanings of words. It's also not appropriate to use for outgoing correspondence as it can result in some pretty bad malapropisms. My favorite is when google translate suggested "fiesta de tarifos" as the translation for "tax holiday." Far from "tax holiday", "fiesta de tarifos" would mean tax party. I'm sure a government official would be bemused reading our inquiry about the tax party.
Re: (Score:2)
My favorite is when google translate suggested "fiesta de tarifos" as the translation for "tax holiday." Far from "tax holiday", "fiesta de tarifos" would mean tax party.
What is the correct translation for "tax holiday?" I would usually expect to see "impuestos" used instead of "tarifos" for taxes...
Re: (Score:2)
Well, here in Oz when 30 June tax year ends and the Accountants have done all their work Tax Holiday should probably translate into "fiesta de tarifos". A tax party.
That is the problem, it is context dependent.
One that I came across is that English has Brother and Sister, where as Indonesian has Kakak and Adik (by foggy memory) for older sibling and yongger sibling. Pretty hard to auto translate.
Re: (Score:2)
The correct translation would be more like "exoneración temporal de impuestos" (literally just "temporary tax exemption"). Calling it a "holiday" is just an English idiom that doesn't have a direct equivalent in Spanish.
"Impuestos" and "tarifos" are basically synonyms. Tarifos is more directly "tariffs", but in some English contexts "tariffs" can be used interchangeably with "taxes." Impuestos would certainly be more common in the Americas.
Re: (Score:2)
Now for some real toughies (Score:1)
Detailed context is required to decide whether 'cleave' is being used to split something apart of join it together, and to decide whaether 'sanction' is permitting something or forbidding it.
Re: (Score:2)
Lawyers like parsley! (Score:1)
Now the AI will think that lawyers like parsley. They said "garnish".
You can't blame AI (Score:2)