News Item: : Language Technologies in online action
(Category: Issues in Linguistics)
Posted by admin
Friday 20 February 2009 - 08:33:29
One of the recent examples of LT tools is visible on SKYNEWS television in the UK. A company produces online captions of what is being said by the people on and off screen. The words of various speakers are marked by different colours and the obvious and hilarious mistakes (misunderstanding of voice input due to incorrect chunking and selection of homophones) re eventually corrected or left as they were, making the reading of the captions a tiresome exercise.
Now the translation of modalities, like speech to text or text to speech makes sense, but the desire to do that online, simulating a kind of human interpreter is a scam. Even if you assume that the program is capable of learning, the underlying philosophy must be to expect a string that the program is familiar with as it has already had a match approved by humans later, the information will be necessarily lost as it is always the new content where it is to be found. (In fact, by misunderstanding the text the program furnishes you with a lot of new, unexpected information which at the end of the day is garbage.)
But this is not the most important issue in this short report. In Machine Translation where a text in L1 is translated into L2 by using Translation Memories similarly to what is seen in the above television news service, you do not translate word for word, or you should not anyway, because what you are trying to do is to find a matching phrase describing a chunk of reality in L1 to make sense in L2 too. Meaning is never complete without context, and the less you know about a subject, the more context will need to be made explicit. There is no such thing as language competence and domain competence separately. If you do not have them both, do not translate anything meant to be a serious text, not a farce or a parody.
But the greatest lesson is not what is outlined above. This footage illustrates the fact that interpreting and/or translating employs a look ahead approach to the job, therefore there is no such thing as knowing a language and knowing your subject matter. Knowledge, skill and competence to translate go hand in hand. It is an anecdotal evidence that once a n interpreter was asked to translate a lecture on birdlife in Hungary and she/he was shown a picture of a cinke (tit). She/he said that it was a bird, then the slide of another songbird vörösbegy (robin) came, and she/he said it is another bird, and so on. In order to be able to translate any text, you need to know your subject inside out. Machine Translation does not have the ways to replace a qualified and professional human translator. Despite building more and more sentence pairs in the Translation Memory, the problem is not solved. Instead, you are creating a software dinosaur, like Windows is one of them already where you do not need 90 percent of the functionalities available. Just think about that.
This news item is from www.firkasz.com
( http://www.firkasz.com/news.php?extend.58 )
(Category: Issues in Linguistics)
Posted by admin
Friday 20 February 2009 - 08:33:29
One of the recent examples of LT tools is visible on SKYNEWS television in the UK. A company produces online captions of what is being said by the people on and off screen. The words of various speakers are marked by different colours and the obvious and hilarious mistakes (misunderstanding of voice input due to incorrect chunking and selection of homophones) re eventually corrected or left as they were, making the reading of the captions a tiresome exercise.
Now the translation of modalities, like speech to text or text to speech makes sense, but the desire to do that online, simulating a kind of human interpreter is a scam. Even if you assume that the program is capable of learning, the underlying philosophy must be to expect a string that the program is familiar with as it has already had a match approved by humans later, the information will be necessarily lost as it is always the new content where it is to be found. (In fact, by misunderstanding the text the program furnishes you with a lot of new, unexpected information which at the end of the day is garbage.)
But this is not the most important issue in this short report. In Machine Translation where a text in L1 is translated into L2 by using Translation Memories similarly to what is seen in the above television news service, you do not translate word for word, or you should not anyway, because what you are trying to do is to find a matching phrase describing a chunk of reality in L1 to make sense in L2 too. Meaning is never complete without context, and the less you know about a subject, the more context will need to be made explicit. There is no such thing as language competence and domain competence separately. If you do not have them both, do not translate anything meant to be a serious text, not a farce or a parody.
But the greatest lesson is not what is outlined above. This footage illustrates the fact that interpreting and/or translating employs a look ahead approach to the job, therefore there is no such thing as knowing a language and knowing your subject matter. Knowledge, skill and competence to translate go hand in hand. It is an anecdotal evidence that once a n interpreter was asked to translate a lecture on birdlife in Hungary and she/he was shown a picture of a cinke (tit). She/he said that it was a bird, then the slide of another songbird vörösbegy (robin) came, and she/he said it is another bird, and so on. In order to be able to translate any text, you need to know your subject inside out. Machine Translation does not have the ways to replace a qualified and professional human translator. Despite building more and more sentence pairs in the Translation Memory, the problem is not solved. Instead, you are creating a software dinosaur, like Windows is one of them already where you do not need 90 percent of the functionalities available. Just think about that.
This news item is from www.firkasz.com
( http://www.firkasz.com/news.php?extend.58 )