![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
Slowly, slowly...
Aug. 1st, 2003 12:47 amA couple of recent news items make it appropriate, I think, to revisit a post from nearly two years ago, on the nature of machine translation.
The first news item was a press release from the University of Southern California, titled Romancing the Rosetta Stone. The piece briefly describes a technique for machine translation that does not rely on the traditional method that, to oversimplify its description, consists of decomposing the source text, analyzing the parts, finding the analogous parts in the target language, and reassembling a target text. Instead, the new technique analyzes the way words stand in relation to one another in existing parallel source and target texts. (I can't help but think of this as a form of neural networking applied to text.)
For me one of the money quotes from the press release is the following:
Almost universally speaking, translators don't think very highly of machine translation, and for good reason. It's because, almost universally speaking, machine translation is not just bad, but spectacularly so, and nearly every translator I know has a personal anecdote to relate in this regard.
My own FMTH (Favorite Machine Translation Howler) occurred when I encountered the following line in a Russian source document, describing the last line of an address in the U.S. (city changed to protect the innocent):
Ye gods.
Somehow, the address line was inserted in the source document, along with some other bloopers, and the end product obviously was never checked afterward.
While disdain for machine translation is nearly universal among working translators, many allow that emotion to extend to the future of machine translation as well. In my opinion, such confidence in the ultimate failure of any attempt to achieve quality machine translation is ill-placed.
Strange as it may seem, I believe it is ill-placed on the basis of what happened to the game of chess.
Back in the late 1940s, the first description of how a computer might be programmed to play chess was published by Claude Shannon. Nobody took the idea seriously, especially since the problem appeared to require resources far in excess of what was commonly available to researchers. In subsequent decades, prototype chess-playing programs were written purely for research purposes (which is to say they all played horrible chess).
In the early 70s, however, some students at MIT entered a program of theirs, named MacHack IV, in an amateur tournament sponsored by the United States Chess Federation. The program scored one draw, after having lost its first five games in a tournament format that, generally speaking, pits players against others with the same score on a round-by-round basis. (In other words, in the sixth round, the program faced a player who likewise had lost the five previous games.)
The program's provisional rating was somewhere in the low 1200s, on a scale that ranges (practically speaking) from about 800 for people who barely know the moves, to 2800 or more for grandmaster-level players. To give you an idea of what a rating difference means, consider that a 400-point advantage over an opponent means you can expect to win against that opponent about 95% of the time, all things being equal.
MacHack IV may have been best-of-breed in the early 70s, but it was far from being in a position to threaten human dominance of the game. Suggestions to the effect that someday, computers might play at a master level were generally greeted with the kind of polite skepticism one generally reserves for people who believe the earth is flat. And the idea of a computer World Champion made flat earth theories sound plausible by comparison.
A couple of years ago, however, a program running on a massively parallel special-purpose computer did beat the World Champion in match play, albeit not for any official title. Moreover, anyone can go down to their local CompUSA and pick up a software package that will consistently beat all but the top few percent of players in the world. The whole process, from the intial tentative description of how it might be done to the triumph of etched silicon and electrical pulses over human wetware took just about 50 years.
But, you say, translation is not chess?
I agree, but the fact of the matter is this: as computer systems become more sophisticated, their ability to undertake tasks of previously mind-boggling complexity increases to the point where the complexity can be tamed. When you combine the taming of complexity with the fact that humans are very adept at achieving goals believed to be impossible in some sort of key way, the result is dynamite. Or a chess-playing computer. Or an airplane.
In this centennial year of aviation, we should keep in mind that many of the best minds of just over a century ago considered heavier-than-air flight to be flat-out impossible. Moreover, the leading lights of the following generation believed the sound barrier to be impregnable, and some of their intellectual heirs were convinced space travel was impossible because the vacuum of space offered nothing to "push against."
In the end, the history of science tells us that the smart money doesn't often win by betting on something being impossible.
So does this mean I believe the systems touted in these articles are the Real McCoy, the killer translation application that will relegate professional translators to the unemployment line? No, not at all. If you read the articles carefully, no claims along those lines are being made, at least not explicitly. The closest anyone came to talking about the quality of their output was the following, from the Times article:
Statistical translation is something to keep one's eye upon. We live in interesting times.
Cheers...
The first news item was a press release from the University of Southern California, titled Romancing the Rosetta Stone. The piece briefly describes a technique for machine translation that does not rely on the traditional method that, to oversimplify its description, consists of decomposing the source text, analyzing the parts, finding the analogous parts in the target language, and reassembling a target text. Instead, the new technique analyzes the way words stand in relation to one another in existing parallel source and target texts. (I can't help but think of this as a form of neural networking applied to text.)
For me one of the money quotes from the press release is the following:
Och is a standout exponent of [this new] method of using computers to translate one language into another that has become more successful in recent years as the ability of computers to handle large bodies of information has grown, and the volume of text and matched translations in digital form has exploded, on (for example) multilingual newspaper or government web sites.The other news story, in the New York Times [Registration required] (From Uzbek to Klingon, the Machine Cracks the Code) also mentions USC and says pretty much the same things, including the following:
[Emphasis mine]
Although in one sense it was more economical, [statistical] machine translation was also much more complex, requiring powerful computers and software that did not exist for most of the 90's.Why did I emphasize the text in italics? Let me back up a bit...
[Emphasis mine]
Almost universally speaking, translators don't think very highly of machine translation, and for good reason. It's because, almost universally speaking, machine translation is not just bad, but spectacularly so, and nearly every translator I know has a personal anecdote to relate in this regard.
My own FMTH (Favorite Machine Translation Howler) occurred when I encountered the following line in a Russian source document, describing the last line of an address in the U.S. (city changed to protect the innocent):
BEVERLY HILLS ПРИБЛИЗИТЕЛЬНО 90210The appearance of ПРИБЛИЗИТЕЛЬНО (which means "approximately") threw me. An address line like that should have said CA, or at worst CALIFORNIA. I was puzzled for a moment, and then it hit me: if you squint real hard, CA could be interpreted as the (capitalized) abbreviated form of 'circa' (that wonderful word that, for example, tour guides in European art museums use when describing dates that are approximate).
Ye gods.
Somehow, the address line was inserted in the source document, along with some other bloopers, and the end product obviously was never checked afterward.
While disdain for machine translation is nearly universal among working translators, many allow that emotion to extend to the future of machine translation as well. In my opinion, such confidence in the ultimate failure of any attempt to achieve quality machine translation is ill-placed.
Strange as it may seem, I believe it is ill-placed on the basis of what happened to the game of chess.
Back in the late 1940s, the first description of how a computer might be programmed to play chess was published by Claude Shannon. Nobody took the idea seriously, especially since the problem appeared to require resources far in excess of what was commonly available to researchers. In subsequent decades, prototype chess-playing programs were written purely for research purposes (which is to say they all played horrible chess).
In the early 70s, however, some students at MIT entered a program of theirs, named MacHack IV, in an amateur tournament sponsored by the United States Chess Federation. The program scored one draw, after having lost its first five games in a tournament format that, generally speaking, pits players against others with the same score on a round-by-round basis. (In other words, in the sixth round, the program faced a player who likewise had lost the five previous games.)
The program's provisional rating was somewhere in the low 1200s, on a scale that ranges (practically speaking) from about 800 for people who barely know the moves, to 2800 or more for grandmaster-level players. To give you an idea of what a rating difference means, consider that a 400-point advantage over an opponent means you can expect to win against that opponent about 95% of the time, all things being equal.
MacHack IV may have been best-of-breed in the early 70s, but it was far from being in a position to threaten human dominance of the game. Suggestions to the effect that someday, computers might play at a master level were generally greeted with the kind of polite skepticism one generally reserves for people who believe the earth is flat. And the idea of a computer World Champion made flat earth theories sound plausible by comparison.
A couple of years ago, however, a program running on a massively parallel special-purpose computer did beat the World Champion in match play, albeit not for any official title. Moreover, anyone can go down to their local CompUSA and pick up a software package that will consistently beat all but the top few percent of players in the world. The whole process, from the intial tentative description of how it might be done to the triumph of etched silicon and electrical pulses over human wetware took just about 50 years.
But, you say, translation is not chess?
I agree, but the fact of the matter is this: as computer systems become more sophisticated, their ability to undertake tasks of previously mind-boggling complexity increases to the point where the complexity can be tamed. When you combine the taming of complexity with the fact that humans are very adept at achieving goals believed to be impossible in some sort of key way, the result is dynamite. Or a chess-playing computer. Or an airplane.
In this centennial year of aviation, we should keep in mind that many of the best minds of just over a century ago considered heavier-than-air flight to be flat-out impossible. Moreover, the leading lights of the following generation believed the sound barrier to be impregnable, and some of their intellectual heirs were convinced space travel was impossible because the vacuum of space offered nothing to "push against."
In the end, the history of science tells us that the smart money doesn't often win by betting on something being impossible.
So does this mean I believe the systems touted in these articles are the Real McCoy, the killer translation application that will relegate professional translators to the unemployment line? No, not at all. If you read the articles carefully, no claims along those lines are being made, at least not explicitly. The closest anyone came to talking about the quality of their output was the following, from the Times article:
Dr. Knight and others said the progress and accuracy of statistical machine translation had recently surpassed that of the traditional machine translation programs used by Web sites like Yahoo and BabelFish.That being the case, translators should have no immediate worries, since Web translations tend to be painfully bad, and nearly useless even for "gisting" (i.e., translations that are admittedly far from perfect, but good enough to give the reader a rough idea, or the gist, of the text's content).
Statistical translation is something to keep one's eye upon. We live in interesting times.
Cheers...
