A few milestones in the short but storied history of machine translation: in 1939, Bell Labs presented the first speech synethesizing device, the Voder, at the World's Fair in New York. In 1978, the first spoken words were transmitted across the Internet. June 2012 saw the release of VoiceTra4U-M, an iPhone app developed by the global Universal Speech Translation Advanced Research Consortium (U-STAR) which enables voice translation of 13 different languages.
Today's translation machines, both written and spoken, "are extremely clever and give us a lot reasons for thought about what language is and how we may understand language better, but the way they work bears little resemblance, in fact, no resemblance at all to the way human beings both speak," says David Bellos, a translator and director of the Program in Translation and Intercultural Communication at Princeton University.
Watch the interview:
Computers decode and reproduce spoken human language in much the same way they translate written language -- by effectively transcribing the speech in the source language into text and putting it through a translation device which "sounds out" the text, "just like your telephone answering device does." (This feature is used and will, says Bellos, always be used in machines that simulate speech translation.) Software translation programs like Google's, Yahoo's, and Microsoft's are essentially statistical engines. Programmers use data to train their algorithms on human-translated parallel texts so that they automatically "learn" how to translate.
Over the years, the technology has become more sophisticated, but speaking to an automated voice on the other end of the line is still an exercise in frustration. The results of programs like Google Translate are notoriously comical. Here, for instance, is Hamlet's famous "To be or not to be" soliliquy translated from the original English to Chinese, back to English again via Google Translate:
Or not, this is a problem:
Whether this is a noble mind suffer
Outrageous slings and arrows of Fortune
Or take up arms against a sea of troubles,
And opposing the closure, after they die, to sleep
A sleep to say we end
The heart of pain, as well as countless other natural shocks
This flesh is heir to it?
As Phil Blunsom, a researcher at Oxford University, told the BBC, "the time when a computer can match the interpretive skills of a professional is 'still a long way off.'"
What's the Significance?
The limitations of machine translation are indicative of the broader historical limitations of symbolic A.I. Early researchers regarded both the human brain and human language as systems of explicit rules which could be pinned down, catalogued, and unlocked -- but despite a few breakthroughs in the field, we've still not come close to building a brain or decoding the nuances of language. Perhaps the problem is more than technological. Perhaps it is unsolvable.
Why? "You possess a skill that hardly any computer programme does," explains the author of a 2009 paper from the University of Copenhagen. In studies, people are able to pick up on subtle distinctions in the meanings of words that computer systems always miss, for example:
(1.1) (a) The command interface deﬁnes a single method called “execute” that is invoked by the
internal CommandExecutor when a command is to be executed.
(b) An Iranian cleric, Hojatoleslam Rahimian, called today for the leaders of Iran’s
opposition Green Movement to be executed.
Google Translate and the automated phone operator fall flat when they try to understand passages that contain complexity and variation -- abstract ideas, shifts in tone, words that mean more than one thing. That's why Bellos believes machine translation will always require the existence of human translators.
Still, he says, machine translation has great potential to expand our sense of the possibilities of communications, as civilization grows increasingly global. "The way airplanes fly resembles not at all the way birds fly. It doesn’t have to. What you want is the flight."
The overall picture is this. The more machine translation there is the more translation will happen, the more people will expect to be able to communicate with other folk and the more they will realize that although machines can clear the ground the actual translation has to be done by somebody because language is human behavior. It’s machine simulated, but they’re not doing anything like what a human translator is doing.
Image courtesy of Shutterstock.
How to Form a Good Habit
The Rules of Power: What Che and Hitler Have In Common
6 people liked this.
Add New Comment
Post as …
Real-time updating is enabled. (Pause)
Showing 10 comments
How do you translate a lie?
2 hours ago 1 Like
HudsonPop Thinker Specializing in Strategy, Systems Architecture & Innovation, Business Troubleshooting and Brainstorming, Computational Linguistics and Intelligence.
Clearly what is needed is software that can naturally perform Language Comprehension. The common place word-for-word transposition (statistical) approach is laughable and will only ever fail.
Although it was a little unfair giving text from Shakespeare to Google Translate since nobody speaks as they did 400 years ago. Always good for a laugh though. Which I think reinforces my first point, that the statistical approach just doesn't work well enough.
To approach a solution, developers need to know a couple of things: 1) Robotics is a silly wasteful way to approach the development of intelligent entities. 2) You don't need super powerful computers to perform the computations needed to achieve comprehension. What we've got right now is more than enough.
As for the question of ambiguity, and this is a tough one for humans to accept, - deal with it. It goes with the intelligence territory. Just as physics has to deal with spooky quantum mechanics, so humans trying to get by in the world have to deal with ambiguity. This means that when you buy your super-duper language comprehending doo-dah it will still suffer the same issues that any human would when it comes to trying to understand what...
15 hours ago
What has become clear to me after a career of teaching medical transcription and transcribing surgical procedures is that this field is not the ideal one for voice recognition software. I am bothered that there is not more concern with the machine bungling such phrases as "a mass the size of an orange" (the final copy reading, "the size of an RN).
And yes, this copy passed through the editor. I believe there is a place for VRS in the workplace, but only if it is accurate. We are still human beings with human needs.
21 hours ago 1 Like
Megan EricksonPop Thinker Megan Erickson is Associate Editor at Big Think, and has worked as a tutor and teacher at several NYC public schools.
Good point--computers are better at catching technical errors, whereas human beings are better suited to interpretation.
0 minutes ago in reply to jbucher
Patrick ElliottPop Thinker
Cyc. There have been projects, like this, which use statistical interpretations, and the ability to process through new words, and old, to find contextual, and other connections. The problem with them end up being a) storage, b) lookup, and c) processing speed. Due to the fact that the means by which words are stored is based on letters, and do use lookups like that, all of these become a problem. In fact, Cyc can be used as OpenCyc, to build a bot, but it a) doesn't learn like the original, b) the data system is a mess to understand, and c) you have to practically program the things understanding of the language from near scratch, due to it not having a way to learn directly, with the result that you might as well be using an Alice bot. It "should" be smarter than that, by a wide margin, but unless you have 5 years to teach it, it never will be, and by the time it is.. it will become unwieldy, slow, and unstable.
Maybe what needs to be done is working on things from the other angle - Have the machine process words as sound, which can...
1 day ago
Computer decoding could develop a database to categorize what words in a particular context refer to--animal, vegetable, mineral, etc.
But this strategy probably would work only in translating texts that are mainly pragmatic and do not rely on metaphors to highlight their points.
I see a basic difficulty in translating language and phrases that have multiple meanings (ambivalent or multivalenced words), such as--
"He ratted to the company chair";
"She's fit as a fiddle."
Decoding would also have to find a way to include word play like puns, slang, and wit.
Idiomatic forms of language don't yield well to mechanical forms of translation.
1 day ago 1 Like
Mr. Zychowski I am the light!
We as humans no longer are able to make machines on our own. Machines are already made by machines, it is not unreasonable to think that a master machine will start coding machines as well. Once that happens the singularity will take over and artificial intelligence will take over the planet. The only thing we need to ponder is weather it will be a good assimilation or bad. Similar to a Start Trek computer or a Terminator. Perhaps it will be both with integration into the biological, some sort of cyborg. These might be the last generations of homo sapiens as we know them. I can't wait to see how it will turn out.
1 day ago
I mean to say that a machine cannot be entitled with the level of capability measured from how he performs with the code given.
1 day ago
The capabilities of machines depend on the capabilities of humans to program. Machines can work better than humans with better code.
1 day ago
Bah to the idea that translation always will require a human. As if there were something magical, or spiritual, or non mechanical in our use of language. The problems and obstacles faced aren't fundamental, it's simply a matter of increasingly complex programming and processing power.
Look at the example given with the word 'execute'. The difference in meaning is in what the word is referring to in each sentence. 'Execute' means something different when it is applied to a person instead of an object like a computer program. Perhaps a program which examines the subject, object, and verb in each sentence and classifies subjects and objects as "person, place, or thing" and verbs as "action, linking, or auxilary" would have more luck interpreting and translating. You could rank words in the sentence on importance, then compare to words in the other language.
The alternative is to create a database of human translated phrases, then once an exhaustive database is created; the machine can choose the appropriate translation from the list. That wouldn't be as hard as one might think. Especially since most humans may know thousands of words and phrases, but only use a fraction of them regularly. For the...
1 day ago
M Subscribe by email S RSS
Show more reactions
blog comments powered by DISQUS
Sign up for our newsletter
Share This Story
About Think Tank
783 Posts since 2010
A place where Big Think's editors and expert contributors look at ideas that are in the news.
The Rules of Power: What Che and Hitler Have In Common
What's Lost (And Found) In Machine Translation
How to Form a Good Habit
Russia in "Survival Mode": Pussy Riot Guilty of "Hooliganism." Garry Kasparov Arrested.
Anxious? Depressed? Literate? Try Bibliotherapy
The Trials of Julian Assange: The End of Diplomacy?
How To Recognize Your Next Brilliant Idea
Update: Hypersonic Aircraft Crashes During Test Flight
70 Percent of Research at Academic Institutions is Wasted
How The Internet Is Killing The Family Recipe
Arts & Culture
Business & Economics
Health & Medicine
Inspiration & Wisdom
Life & Death
Love, Sex, and Happiness
Media & Internet
Politics & Policy
Science & Tech
Truth & Justice
Voted Time Magazine's #1 Website for News and Info
Visit the Floating University, a joint venture between Big Think and the Jack Parker Corporation.
Visit our new online learning course, where Big Think experts deliver actionable career insights. For institutions and individuals.