Are some language combinations better suited for Google Translate? Thread poster: Heinrich Pesch
|
|---|
Heinrich Pesch 
Finland
Local time: 10:41
Member (2003)
Finnish to German
+ ...
I wonder if someone has noticed, that GT performs better for some language combination than for another.
I noticed at least, that when I use GT for translating Spanish into German the output is almost useless, whereas Spanish to Finnish is quite useful. That's unexpected, since Spanish is more closely related to German than to Finnish.
| | | |
| French to English | Nov 4, 2016 |
Heinrich Pesch wrote:
I wonder if someone has noticed, that GT performs better for some language combination than for another.
I noticed at least, that when I use GT for translating Spanish into German the output is almost useless, whereas Spanish to Finnish is quite useful. That's unexpected, since Spanish is more closely related to German than to Finnish.
Is quite good, while French to Italian is not, I also tried out of curiosity Spanish to Italian and found the output really terrible...
| | | |
Lincoln Hui
Hong Kong
Local time: 15:41
Member
Chinese to English
+ ...
| English and European languages | Nov 4, 2016 |
I've found it quite useful when I need a quick translation of the major European languages like Italian, German and French. I've also fed a Norwegian text into Google translate once and gotten a fairly plausible output.
Chinese and Japanese in GT is a complete lost cause at this time, however. It simply doesn't work, whether it's a single word, a short phrase, or an extended text.
| | | |
Lingua 5B
Bosnia and Herzegovina
Local time: 09:41
Member (2009)
English to Croatian
+ ...
| If the results are good.... | Nov 4, 2016 |
that's because someone paid linguists to polish up results, not because languages are related.
I worked on such computational linguistics projects. What is involved is predicting, measuring and fixing errors, doing statistics, etc. Not happening thanks to a magic wand, someone's behind it.
Typically, the pairs they will invest money to optimize and refine results are for languages/markets that have bett... See more that's because someone paid linguists to polish up results, not because languages are related.
I worked on such computational linguistics projects. What is involved is predicting, measuring and fixing errors, doing statistics, etc. Not happening thanks to a magic wand, someone's behind it.
Typically, the pairs they will invest money to optimize and refine results are for languages/markets that have better economic value/purchase power. Linguists are paid hourly, but also the programmers must be paid. It costs money. ▲ Collapse
| | |
|
|
|
Heinrich Pesch
Finland
Local time: 10:41
Member (2003)
Finnish to German
+ ...
TOPIC STARTER | No-one polishes here | Nov 4, 2016 |
Lingua 5B wrote: that's because someone paid linguists to polish up results, not because languages are related. I worked on such computational linguistics projects. What is involved is predicting, measuring and fixing errors, doing statistics, etc. Not happening thanks to a magic wand, someone's behind it. Typically, the pairs they will invest money to optimize and refine results are for languages/markets that have better economic value/purchase power. Linguists are paid hourly, but also the programmers must be paid. It costs money.
I was talking on text output I generate myself on the fly, no paid jobs.
| | | |
Lincoln Hui
Hong Kong
Local time: 15:41
Member
Chinese to English
+ ...
| Different thing | Nov 4, 2016 |
She's saying that GT's output is reviewed and optimized by humans who tweak the programming and library to make it produce better translations.
| | | |
Lingua 5B
Bosnia and Herzegovina
Local time: 09:41
Member (2009)
English to Croatian
+ ...
Lincoln Hui wrote:
She's saying that GT's output is reviewed and optimized by humans who tweak the programming and library to make it produce better translations.
That's what I'm saying. I'm surprised nobody invested in Chinese output, it's a big language. It's usually the target language country that's funding it. Perhaps they saw no sense in it.
| | | |
Michael Beijer
United Kingdom
Local time: 08:41
Member (2009)
Dutch to English
+ ...
| Google Translate can also be very good at Dutch into English, | Nov 4, 2016 |
as is Microsoft Translator, and to a lesser extent SDL's own Language Cloud offering (which will soon be getting the new AdaptiveMT system in Studio 2017, which I'm now testing).
Not sure I agree with Lingua 5B's comments on the output being "polished up" by linguists. I'd have to see proof of that before I'd believe it, as it's not sth I've heard before.
[Edited at 2016-11-04 09:31 GMT]
| | |
|
|
|
Sheila Wilson
Spain
Local time: 08:41
Member (2007)
English
+ ...
| My personal (not professional) experience | Nov 4, 2016 |
I use GT in my personal life to translate long Spanish texts into English as I can't (yet) understand them in Spanish. I can generally get the sense of a paragraph but a page calls for so much time and concentration I feel defeated from the first word. Sometimes, GT provides reasonable output that I can interpret - it's rarely in correct English but then I specialise in polishing English  . But quite often it's unintelligible. And ... See more I use GT in my personal life to translate long Spanish texts into English as I can't (yet) understand them in Spanish. I can generally get the sense of a paragraph but a page calls for so much time and concentration I feel defeated from the first word. Sometimes, GT provides reasonable output that I can interpret - it's rarely in correct English but then I specialise in polishing English . But quite often it's unintelligible. And for some reason it seems blind to the little Spanish word "no", resulting in negative verbs turning into affirmatives. I'm really surprised that GT hasn't made a little more progress in such a common pair.
Just as a translation aid when operating outside my comfort zone, I once put a bog-standard French partnership agreement through GT. The English output was extraordinarily good. Very worrying, actually  . It made me thankful that I specialise in marketing texts. ▲ Collapse
| | | |
| What about between Dutch and French? | Nov 4, 2016 |
Based on the theory that there would be an extremely high proportion of high-quality translations on account of Belgium and relatively few low-quality translations for an otherwise improbable pair of languages ...
I was curious and just tried two paragraphs of a non-descript newspaper article from the "Berliner Zeitung" (93 words in 6 sentences) for German > English. If you can understand German, it is generally easy to guess what the original probably meant. There is one sentence t... See more Based on the theory that there would be an extremely high proportion of high-quality translations on account of Belgium and relatively few low-quality translations for an otherwise improbable pair of languages ...
I was curious and just tried two paragraphs of a non-descript newspaper article from the "Berliner Zeitung" (93 words in 6 sentences) for German > English. If you can understand German, it is generally easy to guess what the original probably meant. There is one sentence that is largely misleading (the verb is assigned to the wrong subject). All of the sentences are somewhere on the threshold between being gibberish and being puzzling, but essentially comprehnsible. It's hard for me to guess how much someone who can't understand German would get out of this translation.
I did the same thing with a 148-word sample from the "New York Times" (English > German). There is a schockingly convincing bit at the beginning, which was probably caused by a long and more or less standardized phrase, but then the Google Translation descends into more or less complete gibberish, including major terminological problems (possible but obviously implausible choices for two central terms).
I would think that MT generally has trouble working out of English, because individual words in English tend to have so many meanings (compared to German, at least) and because English grammar does so little to formally indicate parts of speech (with the exception of adverbs, most words could theoretically play several roles ... for example, "yellow" is a noun, adjective, and verb). Does this theory hold up?
I also assume the subject matter and text type matter a lot. ▲ Collapse
| | | |
| German - Dutch / Russian - Dutch / Hebrew - Dutch | Nov 4, 2016 |
German-Dutch can be suprisingly good for certain topics. Looks like a lot of technical and chemical terminology is fed into the system. Of course it's still far from publishable, but at least you understand what the text is about. Russian - Dutch and Hebrew - Dutch are simply horrible and useless.
| | | |
Lingua 5B
Bosnia and Herzegovina
Local time: 09:41
Member (2009)
English to Croatian
+ ...
| Confidentiality. | Nov 4, 2016 |
Michael Joseph Wdowiak Beijer wrote:
as is Microsoft Translator, and to a lesser extent SDL's own Language Cloud offering (which will soon be getting the new AdaptiveMT system in Studio 2017, which I'm now testing).
Not sure I agree with Lingua 5B's comments on the output being "polished up" by linguists. I'd have to see proof of that before I'd believe it, as it's not sth I've heard before.
[Edited at 2016-11-04 09:31 GMT]
How can I show you proof if I'm working under NDA?
Why would they be measuring results and ask reports to be made about results, just for fun? If this process is not done for a language pair, the results will be statistically worse (ie. likely worse). The more layers and repeated checks, will minimize errors even further. It is more important for them in certain language pairs than other ones, mainly for trade and e-trade reasons.
If someone sends you a project to analyse output words, phrases and sentences, and dynamic of errors, what would be the final purpose of this being done? If you do this process for a language pair 1k times, and for another language pair just 10 times, what pair would statistically produce better results?
| | |
|
|
|
Heinrich Pesch
Finland
Local time: 10:41
Member (2003)
Finnish to German
+ ...
TOPIC STARTER
But I can hardly believe that someone would optimize GT for a minor language like Finnish, with only 5 million speakers, and not for German.
For the Spanish into German pair there seems to be some fixed rules that make for funny results. All persons speaking to each other address the other in the 3rd person plural, even a father speaking to his child. Even though there is in Spanish "usted" for the purpose, all other cases should rather be translated into 2nd person singular.
Also G... See more But I can hardly believe that someone would optimize GT for a minor language like Finnish, with only 5 million speakers, and not for German.
For the Spanish into German pair there seems to be some fixed rules that make for funny results. All persons speaking to each other address the other in the 3rd person plural, even a father speaking to his child. Even though there is in Spanish "usted" for the purpose, all other cases should rather be translated into 2nd person singular.
Also German to Finnish in GT delivers often acceptable results.s at least for technical and scientific purposes. ▲ Collapse
| | | |
Login to reply/comment 
Translation news
