If it makes those up north feel better, computer translation was next to useless for anything slightly complex until Hinton and friends figured out deep learning.
Yes, I remember using a translation startup's automated website to translate a text to French where "kind regards" were translated to "I'd like to thank the president of the EU". I'm pretty sure they were using EU documents for training...
I believe other sources include UN documents, which are translated into even more languages, and then there are a bunch of combined corpora like https://opus.nlpl.eu/ which contains hundreds of languages as well.
The obvious downside of this, of course, is that translated text often ends up quite official and stiff. Great for reading news, translating speeches and reading papers, not so great for day to day conversation. There are corpora for more practical, specific use for a few languages, but they usually only come in pairs.
I don't know how services like Google Translate work now, but through experimentation you used to be able to see how the translation matrix went through English first when you translated between two other European languages. You'd put in a word for which the source and destination language had specific words, but for which English only had one single word, and you'd get a mistranslation every time. The shorter the sentences, the more likely the translation service would get it wrong.
I just found out this can still be replicated on Google Translate, actually: the Dutch word mees (tit, the bird, aka titmouse) and the Dutch word tiet (tit, slang for human anatomy) both get translated into French as mésange, the bird, when you enter them each on a new line. Add more words for context ("wat een grote rieten" vs "wat een grote mezen") and the translation corrects itself.
I don't believe the French word mésange has any direct correlation with mammal anatomy, so clearly the translation matrix involves the English language, or af least one with similar ambiguities. This is why you can't rely on Google Translate if you're just looking for the translation of a single word, especially if you're translating between languages other than English.
To your point, Google translate still has issues translating words for every day objects between Danish and English. Probably the same issue arrises for most smaller languages with less parallel text.
Just yesterday I wanted to know what an "urtekniv" is called in English. Google translate just returned the word untranslated. Ironically my second try was to google "urtekniv engelsk" and I got an answer from a dictionary: https://www.wordsense.eu/urtekniv/
Maybe Google Translate should accept they don't have the data to learn stuff like this naturally, and just inject some dictionaries into the mix.
I've found machine translation often does a poor job even with news, in European languages. There's just too much implicit shared context we don't even think about.
The mistranslated example I remember best is the word "Bund" in a German headline, which can mean a lot of things, but in context clearly meant the federal government.
Also of note, a number of pioneering studies pertaining to the effect of bilingualism on cognition and cognitive development were undertaken in Canada as well.
The norm is multilingualism across different languages that are all local to the area. French people studying English in school are not "bilingual" at all, in this sense.
We have tried to raise our daughter at least with some bilingualism, and so had her in French immersion from senior kindergarten through to grade 8. In retrospect a mistake for a multitude of reasons. I doubt her French will persist well beyond her school years and without a community other than school to use it in, it's mainly pointless, and many aspects of the curriculum were harder than they should have been for her because of the language barrier. And after 9 years of immersive education she can't listen to a native French speaker and understand or watch TV, because French in the real world and outside of her classroom is spoken much more quickly and accented.
Nevermind that French immersion in Ontario is teaching these kids Parisian French and pronounciation, while 25% of the population of Canada speaks Quebecois dialect so kids here are not learning the French from their own country. Frankly, there's a bias in the education system against Quebecois French generally which reflects a Colonial era mentality (mother country "good" -- and local is a poor imitation, a psychological disease that's all over Canada, especially English Canada).
All of this ranting is just to bring up the challenges of trying to attain bilingualism when you are a native English speaker in an English speaking country. In some ways it's a curse.
These days I'd mostly prefer to live in Quebec for cultural/political/lifestyle reasons. But I was brought up unilingual in Alberta (with some German), and never learned French beyond half-heartedly counting to 10.
Canada's official bilingualism is mostly a thin veneer that is a reality on the ground only in places like Ottawa. Sort of a half-hearted half-completed project just to try to put a cork in separatism.
It's strange that in the middle of the article, there's suddenly a paragraph about how government requirements on research and procurement usually are detrimental. It's like they're saying "this story is neat, but let's not forget government is awful." At least finish your story before criticizing unrelated things.
How is it not topical to mention that a government-led research project failed to accomplish with their own data trove what a private organization did with it, when the whole article is about that very data and that private organization's achievement?
As someone who had worked with both research and procurement agencies in the Canadian government, I'm surprised the author was able to hold their tongue until the middle of the article.
As someone who has watched a few parliamentary sessions, I smile at the thought of training computers on those recordings. Ever look at a Markov Chain output and think "those are paragraphs of text saying nothing at all"? Well, maybe it's closer to the humans they're emulating than we thought.
I thought translation took a leap forward when Google or someone figured out to use UN documents which are carefully translated from one language to another?
I'm baffled that dense contracts with so much money on the line do not have better clarification of the terms. It would take one or two extra sentences to make intentions clear for both parties (e.g. "For example, if neither party notifies the other by the end of the 4th year that they do not intend to cancel the renewal, the contract will automatically renew for the subsequent 5 years").
I'm not a lawyer, but when I was helping write agreements for my dad with tenants he had - it seemed like a no-brainer to include "this means, if X does not happen by 11:59pm of Y date, Z will apply".
ps - I'm angry at any contract agreement where the date can be interpreted as "end of day, 11:59pm" or "the day, which starts at 0:01am"
How many bugs in software result from forgetting to handle an edge case?
Not surprising the same thing can happen in contracts; they're both written by humans. And you can't unit test a contract! (ignoring smart contracts for now)
I took a philosophy course in my university focusing on law, covering contracts. The professor gave an example "a band wants to play at a venue, they make a contract; what happens if the venue burns down, who owes who money?"
There are definitely edge cases, but it doesn't feel too hard to specify a "catch all", and if the other party doesn't like it, they should clarify those terms too. Feels easy enough. Reminds me of "burden of proof tennis" - if the catch-all isn't satisfactory to you, carve out an exception the other party accepts.
In Ireland (ROI), Irish being the first official language of the state, constitutionally the Irish language version of a law takes precedence (Article 25, 4.6º) over an English language version (this applies to the constitution itself also).
Many laws are written in English in the first place though (the president can sign either form), sadly it's sometimes a decade or more before an official translation is made. A secondary problem is the insistence of a defendant to have his trial heard in the Irish language, or have Irish forms of the various laws/summons/transgressions. We usually call those people "cute hoors".
Back in the 80's, researchers also worked on the concept of a "bridge" language; that is, first translating a text into a language with a simple grammar, few homonyms and other beneficial traits, and then translating again into the target language. One of the bridge languages was Aymara, but then Google's statistical model took over. You can almost tell how much material Google has to work with. For example, Mandarin<=>English translations are pretty good, while Thai<=>English is usually unintelligible.
35 comments
[ 3.4 ms ] story [ 96.3 ms ] threadI believe other sources include UN documents, which are translated into even more languages, and then there are a bunch of combined corpora like https://opus.nlpl.eu/ which contains hundreds of languages as well.
The obvious downside of this, of course, is that translated text often ends up quite official and stiff. Great for reading news, translating speeches and reading papers, not so great for day to day conversation. There are corpora for more practical, specific use for a few languages, but they usually only come in pairs.
I don't know how services like Google Translate work now, but through experimentation you used to be able to see how the translation matrix went through English first when you translated between two other European languages. You'd put in a word for which the source and destination language had specific words, but for which English only had one single word, and you'd get a mistranslation every time. The shorter the sentences, the more likely the translation service would get it wrong.
I just found out this can still be replicated on Google Translate, actually: the Dutch word mees (tit, the bird, aka titmouse) and the Dutch word tiet (tit, slang for human anatomy) both get translated into French as mésange, the bird, when you enter them each on a new line. Add more words for context ("wat een grote rieten" vs "wat een grote mezen") and the translation corrects itself.
I don't believe the French word mésange has any direct correlation with mammal anatomy, so clearly the translation matrix involves the English language, or af least one with similar ambiguities. This is why you can't rely on Google Translate if you're just looking for the translation of a single word, especially if you're translating between languages other than English.
Just yesterday I wanted to know what an "urtekniv" is called in English. Google translate just returned the word untranslated. Ironically my second try was to google "urtekniv engelsk" and I got an answer from a dictionary: https://www.wordsense.eu/urtekniv/
Maybe Google Translate should accept they don't have the data to learn stuff like this naturally, and just inject some dictionaries into the mix.
I've found machine translation often does a poor job even with news, in European languages. There's just too much implicit shared context we don't even think about.
The mistranslated example I remember best is the word "Bund" in a German headline, which can mean a lot of things, but in context clearly meant the federal government.
Nevermind that French immersion in Ontario is teaching these kids Parisian French and pronounciation, while 25% of the population of Canada speaks Quebecois dialect so kids here are not learning the French from their own country. Frankly, there's a bias in the education system against Quebecois French generally which reflects a Colonial era mentality (mother country "good" -- and local is a poor imitation, a psychological disease that's all over Canada, especially English Canada).
All of this ranting is just to bring up the challenges of trying to attain bilingualism when you are a native English speaker in an English speaking country. In some ways it's a curse.
These days I'd mostly prefer to live in Quebec for cultural/political/lifestyle reasons. But I was brought up unilingual in Alberta (with some German), and never learned French beyond half-heartedly counting to 10.
Canada's official bilingualism is mostly a thin veneer that is a reality on the ground only in places like Ottawa. Sort of a half-hearted half-completed project just to try to put a cork in separatism.
https://www.nytimes.com/2006/10/25/business/worldbusiness/25...
Whenever I’m looking for a legal loophole, I’ll take a look at the French translation of the law.
I'm not a lawyer, but when I was helping write agreements for my dad with tenants he had - it seemed like a no-brainer to include "this means, if X does not happen by 11:59pm of Y date, Z will apply".
ps - I'm angry at any contract agreement where the date can be interpreted as "end of day, 11:59pm" or "the day, which starts at 0:01am"
Not surprising the same thing can happen in contracts; they're both written by humans. And you can't unit test a contract! (ignoring smart contracts for now)
There are definitely edge cases, but it doesn't feel too hard to specify a "catch all", and if the other party doesn't like it, they should clarify those terms too. Feels easy enough. Reminds me of "burden of proof tennis" - if the catch-all isn't satisfactory to you, carve out an exception the other party accepts.
Many laws are written in English in the first place though (the president can sign either form), sadly it's sometimes a decade or more before an official translation is made. A secondary problem is the insistence of a defendant to have his trial heard in the Irish language, or have Irish forms of the various laws/summons/transgressions. We usually call those people "cute hoors".