Coming Soon, The Google Translator

compuglot writes "Google gave journalists a glimpse of its next generation machine translation system at a May 19th Google Factory Tour. "Google Blogoscoped" offers an excellent overview of the presentation. The system has been trained using the United Nations Documents as a corpus. This corpus is some 20 billion words worth of content. It uses existing source and target language translations (done by human translators at the U.N.) to find patterns it then uses to build rules for translating between those languages. Apparently it was successful where the current version had failed in translating certain phrases. If anyone were capable of making a serious go of MT, that would have to be Google."

fascinating

since the RTFAs lacked any kind of crunchiness, i sourced some great stuff here [jhu.edu] that does a wonderful job explaining how this system works, and gives the advantages the statistical translation method has over the rules-based approach. as well as the disadvantages.

fascinating stuff:

"Currently, most machine translation technology, including consumer-oriented programs such as Systran's Babel Fish, have been "taught" the rules of language, such as verb tenses and when to use parts of speech. Programmers painstakingly hand-build systems based on such rules. "The computer is told, if you see this thing in Russian, replace it with this thing in English," explains Yarowsky.

"While somewhat effective, such systems are time-consuming to build (consider how long it takes most humans to learn a language and all its rules), and resulting translations are still marred by grammatical and other errors. Those that do work fairly well usually tackle popular Western languages, such as French, German, and Spanish; there are few translation programs developed for other important tongues, such as Chinese, Turkish, or Arabic, let alone for more obscure languages like Tajik.

"To tackle a broader range of the world's languages, and to improve on the quality of machine translation, Yarowsky and his Hopkins colleagues are developing computer programs that can be trained to figure out any language using statistical analysis, i.e., looking at the probabilities of language patterns. In what's known as automatic knowledge acquisition, the computer could "learn" Serbian well enough to translate future documents or conversation, or at the least pick out pertinent words like "bomb."

"As Yarowsky explains: "Say you want to teach a computer how to translate Chinese: You give the computer 100,000 sentences in English and the same 100,000 sentences in Chinese and run a program that can figure out which words go to which words. If in 2,000 sentences you have the word Washington, and in about the same number of sentences you have the word Huashengdun, and they occur in the same place in the sentence, these words are likely translations.

"It's all just observation," Yarowsky adds. "Children do the same thing, but they also do it through visual stimulation and feedback. They see a book and hear the word 'book,' and eventually they learn that it's a book. They see a bird with its wings flapping around and learn that is called a bird. It's the same with machines, only they have much better memories. Computers could remember exactly when and where they saw the words bird and book."

"So, instead of telling a computer how to do something -- conjugate the verb 'to be' in Spanish, for example (I am = soy) -- researchers give it tens of thousands of examples and program the computer to find repeated patterns that the computer can use to conjugate new verbs. Trained this way, the program could potentially "learn" phrase structure and the rules of translation.

"As Yarowsky notes in his 100,000-sentence example, one way to accomplish automatic knowledge acquisition is to use bilingual or parallel text. The program "reads" a document in English and then a version in a second language. Such texts used by Hopkins researchers include the Bible, which is available on the Web in more than 60 languages, the Book of Mormon (over 60 languages), and the United Nations Declaration of Human Rights (240 languages).

"Aiding the computer is the fact that the English version of such texts can be annotated by hand or using another computer program -- essentially marked up to show, for example, that Jesus is a noun and pray is a verb. The translation program-in-training needs such information because it cannot translate future text just by substituting individual words in each language; it must also be able to analyze how sentences work. To do so, the computer program uses pattern recognition templates and other tools to understand sente

Re:fascinating

You go all the way to Iran to get gasoline? Who are you, George W Bush?

Re:fascinating

Some questions:

Why can't a dictionary be made of nouns, of verbs? Why can't we have it statistically analyze the grammar for ambiguous words?

Does it only recognize exact matches? Especially with verb conjugation, I'd think any words 80% similar or so should be considered matches. Not all languages are as conjugation happy as latin or spanish or even english, and you often lose some nuanced conjugations when translating from one to the other.

What will be done about idioms? Translating these word for word often makes no sense at all, and for me at least (no idea what the official stance is), I'd rather they substitute in idioms with the same general meaning, but for the culture being translated to.

Does it work on alternate character systems, is it word boundary dependent?

Does it understand punctuation rules, will this post translated to spanish have the upside down question marks where they're supposed to be?

How many of the world's existing languages have enough text for this to even be feasible?

Re:fascinating

What will be done about idioms? Translating these word for word often makes no sense at all...

The often-quoted examples are: "Out of sight, out of mind" becomes "invisible idiot" and "the spirit is willing, but the flesh is weak" comes out as "The meat is rotten, but the wine's great".

How many of the world's existing languages have enough text for this to even be feasible?

Ah yes, that's the tricky part. Translating for preservation near-extinct languages that are in spoken or recorded form only. A true programming challenge.

I find the Babel-Fish translator to be nearly useless and the Systran box at www.systransoft.com very helpful when selling things on eBay to people in non-English-speaking countries. When I get a question about an auction item that has little grammar cohesion and has a offshore domain, like
"How many cost you Italia he transport?", I'll run my response through Systran's translator and add the original english afterwards. More often than not the sales and PayPal transactions are successful.

I believe that machine translation will be the 'killer application' for 64-bit home PCs. ..along with DRM busting..

There are five levels of machine translation:

1) word substitution.
2) phrase substitution.
3) cohesive paragraphs and idioms.
4) light literature, magazine articles, and business.
5) classical literature, law, and diplomacy.

Each level requires at least an order of magnitude more computing power than the previous one. Babel fish is on level two and systran is on three. Google is positioning themselves to be between levels four and five.

I wish them the best of luck. Without sarcasm or irony. This is important work.

"Give me a one sentence definition of 'irony'."
"Yeah, it's where the Iranians come from."

Re:fascinating

What will be done about idioms? Translating these word for word often makes no sense at all, and for me at least (no idea what the official stance is), I'd rather they substitute in idioms with the same general meaning, but for the culture being translated to.

I think this is precisely where statistical approaches can really shine. A purely dictionary-based conversion will translate an idiom word-for-word, which will make no sense at all. However, a statistical approach could be constructed to look for the "longest reliable match." So if the idiom "cat got your tongue" re-appears over and over, and is correlated to a different idiom in other languages (that may not use the word "cat"!), then the algorithm could tokenize "cat got your tongue" as a single entry that would map to something different in each language.

How many of the world's existing languages have enough text for this to even be feasible?

You're right... that's the killer. Translating using statistics (especially idioms) properly will require a huge database of samples. Even what's been suggested so far is not enough. If we want to translate technical documents, we need a new database. If we want to translate "free form writing" we need yet more data.

However, there's lots of data out there (already in digital format) that could be used... we just need people to see the potential and start using these datasets (or making these datasets available). For instance, for technical stuff there are thousands of abstracts for papers and for theses that are translated into various languages (for instance, many articles published in german are then also released in english... I live in Quebec, and every thesis abstract has to be translated into french also... etc.). Many legal documents (many of which are already available to the public) are also translated for various reasons. It would also be interesting if translators all around the world uploaded documents they had translated into some database (assuming it's nothing sensitive of course!). As this database grew, it would become more and more reliable. Let's face it, there's tons of human-based translation going on, forming a massive dataset... but by and large it's just scattered and not useable.

Re:fascinating

Don't forget that many works of fiction are translated into several languages. The only problem with that is persuading the copyright holders to permit their use in training computer translation systems. I'm not sure where you would stand with this legally (After all, IANAL!), so I suspect this is why Google has been using the UN documents. I would imagine these are effectively public domain; and if not, I would imagine the UN would see a reliable machine translation project worth supporting. The only downside I can see is that the UN texts are unlikely to have many idioms or colloqualisms, which would limit the resulting translators usefulness in a more general context.

Re:fascinating

Sounds like a very good approach, but am I the only one to see an issue in the texts they're using that are already available in multiple languages?

The examples given (two religious texts and a legal one) don't really sound like the best things for teaching a "blank slate" program a new language. I understand that it's looking for structure and rules rather than word-for-word links, but the Bible uses many outdated or non-standard phrases and sentence structures, as does most legal text I've ever seen. I'm not a linguist or a statistician, but from my uneducated viewpoint it sounds like problems might arise in the texts that are available for training the system. Anyone know how they're planning to overcome this?

Re:fascinating

Well the bible is hebrew, greek and latin. There are no outdated English phrases in the Bible. Now if your refering to the King James translation of the bible, obviously such would be good for teaching google Old English but not modern english. You would need a much newer translation that doesn't use old phrases. Such do exist btw.

Re:translations of translations

I was wrong about the french. However the spanish NVI appears to parallel the NIV, and i'd imagine would be pretty good candidates for this sort of analysis.

http://www.booksofthebible.com/p2390.html [booksofthebible.com]

I believe it's key that in the situation of

Ancient Lang A -> Modern Lang B -> Modern Lang C

that B and C will be far closer than

Ancient Lang A -> Modern Lang B
Ancient Lang A -> Modern Lang C

Re:fascinating

but the Bible uses many outdated or non-standard phrases and sentence structures, as does most legal text I've ever seen. I'm not a linguist or a statistician, but from my uneducated viewpoint it sounds like problems might arise in the texts that are available for training the system. Anyone know how they're planning to overcome this?

Harry Potter is the answer. It is several "normal language" books and is translated to all major languages. Also, program would finally figure out how to translate words like "Quidditch".

Re:fascinating

or at the least pick out pertinent words like "bomb."

Why do I have a funny feeling that this research isn't being funded by philanthropic foundations?

-Eric

Re:fascinating

"Open the pod bay doors, HAL."

"STFU, Dave. LOL!"

except, no.

"It's all just observation," Yarowsky adds. "Children do the same thing, but they also do it through visual stimulation and feedback. They see a book and hear the word 'book,' and eventually they learn that it's a book. They see a bird with its wings flapping around and learn that is called a bird. It's the same with machines, only they have much better memories. Computers could remember exactly when and where they saw the words bird and book."

Except, no. Humans are basically generalization machines. Babies are able to grasp very quickly that words apply to categories of things -- not just that a *specific* item is a bird or a book, but to learn "I know a bird when I see it", even without necessarily being able to provide a scientific definition. Computers can be built to emulate this ability, but learning word-to-word mappings isn't *nearly* the same as learning abstract concepts and which words apply to them.

Needs a *bit* more work...

Just to illustrate, here's the summary of this story, translated to German and back to English using Google's current version [google.com]:

Google gave a Glimpse of its machine Uebersetzungsystems the following production at the factory route of the A May 19 to journalists. Google. "Google Blogoscoped" offers an excellent overview of the representation. The system was trained with the nation documents as korpus. This korpus is something 20 billion word value of contents. It uses the existing target language translations (takes place via human translators at the U.N.) Samples find, which use it then to establish guidelines for translating between those languages. Apparent it was successful, where the present version had failed, if it translated certain cliches. If everyone of forming a serious were capable, of the M.Ue., those would go to have having to Google.

Bork bork bork!

Here is the result as interpreted by the Swedish Chef:

"Guugle-a gefe-a a Gleempse-a ooff its mecheene-a Uebersetzoongsystems zee fullooeeng prudoocshun et zee fectury ruoote-a ooff zee A Mey 19 tu juoorneleests. Guugle-a. "Guugle-a Bluguscuped" ooffffers un ixcellent ooferfeeoo ooff zee representeshun. Zee system ves treeened veet zee neshun ducooments es kurpoos. Thees kurpoos is sumetheeng 20 beelliun vurd felooe-a ooff cuntents. It uses zee ixeesting terget lungooege-a trunsleshuns (tekes plece-a feea hoomun trunsleturs et zee U.N.) Semples feend, vheech use-a it zeen tu istebleesh gooeedelines fur trunsleteeng betveee thuse-a lungooeges. Epperent it ves sooccessffool, vhere-a zee present ferseeun hed feeeled, iff it trunsleted certeeen cleeches. Iff iferyune-a ooff furmeeng a sereeuoos vere-a cepeble-a, ooff zee M.Ue-a., thuse-a vuoold gu tu hefe-a hefeeng tu Guugle-a."

Looking forward to a www.borkle.com which returns all its results in such a format.

Google's translator

So what powers Google's current translator? I have seen it give word-for-word the same as Babel on some occasions (but with better handling of non-ASCII characters).

Re:Google's translator

SystranSoft's Systran [systransoft.com] is behind almost all of the machine translation srevices on the Internet, lincluding Google's.

Integrate with GMAIL!

Make this work with Gmail and I'd even pay money for it!

Tired of getting email from Amazon.DE on my Gmail account and having to copy and paste it over to Babelfish.

That would be very useful for me.

Anyone care to make a bet?

That Microsoft will announce a new revolutionary language translation service sometime in the next two weeks or so?

Re:Unsupported assertions

No, it's because Google has tons of talent, money, already-archived text to work with, computers, respect in the industry, and consumer base. I can't think of a company that possesses these characteristics more so than Google.

Re:Unsupported assertions

Dell and Cisco are not in this business. IBM is not hemorrhaging with cash in the way Google is. Microsoft is not in the business of providing free Internet accessories. In any case, Google has a track record of innovative ideas ("innovative ideas" meaning that not only did they come up with it and implement it partially, but they invested full-on into it, bet money on it, and made it better than the competition) and is most likely of any company who would announce this to actually pull through with it. If some little start-up announced this (as I'm sure a few have), people would take it with a grain of salt. But that Google announces it, I'm sure most people believe fully that Google will deliver on its promise.

And you're right, people have thought of this exact idea (I'm sure every other computer major and linguist has, in fact, since the birth of ENIAC--I know the idea's crossed my mind tons of times, not that I'd have the slightest clue how to do it), however actually attempting to do it with a reasonable chance of success? I'm going to say Google is the first.

Plus, I got the impression from the article that the serve is operational, just not available to the public. If you'll read the article, you'll find that the translator properly translated a fairly complicated phrase from Arabic to English. I'd guess that this service is, from a technical standpoint, at least 95% done -it's just the packaging and touching-up that needs to be done.

Re:Unsupported assertions

Ummm, geeks like Google because Google employs scientists. Which mere scientists were you talking about?

Were you talking about the PhDs at universities busy teaching classes, churning out research papers to avoid being fired (an ugly numbers game some departments play), or perhaps burning time generating volumes of grant paperwork?

Oh, maybe you were talking about the scientists employed by the private sector. I'm sure the management teams wherever they work are willing to take the time and care that Google won't.

You do know how may PhDs Google employs, right? Not to mention that they won't be fighting for resources there either. No backstabbing, liquidating MBAs trashing their corporate budget. No football-crazed alumni assassinating their funding proposals either.

Also, I would remind you that "mere scientists" often come up with the needed research (there are volumes in MT alone), but rarely can afford to put in the years that it takes into a good implementation.

Geeks love Google because it is, in many respects, where the best of business meets the best of academia.

Piffle

If anyone were capable of making a serious go of MT, that would have to be Google.

An interesting story, but please, for the love of all that's balanced and objective; tell me again how that smudge on your nose really is chocolate.

Altavista Babelfish

When questioned on the matter, Altavista's Babelfish translator gave this quote:

Google does not have anything on my amazing abilities of the translation!

if anyone...

Actually, my bet for most likely to make a real go of machine translation would be...

IBM

Look how far they ran with chess programs, because they felt like it...

If they decided to go the same distance with translation...

Re:if anyone...

Yeah right. Not while they're trying to convince customers to buy their current generation of crap translators. I got sucked into an IBM conference two years ago where they tried to convince me that their Websphere translator was "near perfect" and that it was ready to be deployed on web sites wanting to offer content in multiple languages. They even went so far as to bring in supposed unbiased happy customers who testified that the Websphere translator was as good as human translators.

In the conference was mostly IBM platinum partners (development firms who specialize in IBM "solutions" and make IBM enough money to be called platinum partners) and they seemed to buy into it. Of course, platinum partners tend to believe everything IBM tells them.

Only works for translating speeches

If your blog sounds like a politician giving a speech at the UN, this service will do a wonderful job. Doubtful that it will do any better that Babelfish otherwise.

The biggest problem in artificial intelligence is that the system learns the material that it is trained to, and only that material. Computers don't generalize or extrapolate the known into the unknown worth a damn.

Wait, why?

"Computers don't generalize or extrapolate the known into the unknown worth a damn."

Fortunately, that's not all that google has to go on. Google has 8 billion webpages, in many different languages, most of which are written by non-speechwriters. Not only can they analyze words based on translated context, but they can analyze words based on intra-language context, to form associations between words and meanings.

The real trick is getting down two important linguistic concepts: "Sandhi Rules" (for instance, the use of "an" before a vowel and "a" before a consonant, which are totally regular but more complicated than a word-to-word matchup), and the "degree" or "quality" of words, which indicate the type of adjective most appropriate in any given context.

For instance, "erudite", "learned", "educated", "knowledgeable", "skilled", and "cunning" could all be related words, but many of them have positive or negative assocations which may only really be conveyed by understanding the meaning, irony, or sarcasm of a particular phrase.

For instance, "John has been skilled in writing beautiful code for most of his adult life" is quite different from "John has been educated in writing beautiful code for most of his adult life", or "John has been erudite...". The first one is probably right if John has had a natural inclination to doing it properly, the second if he has undergone some training (though we don't know the actual state of his ability), the third (though the word doesn't even really make sense here) if he has been arrogant about his ability, shouting RTFM! every time someone asked him a question.

Good online translators for other languages

While Google's existing translator and Altavista's Babelfish are good, they do not help in the translation of several other languages.

That would be a really good benefit - for instance, I wanted something translated to and fro from Svensk (Swedish), but I really couldn't find any translation service that did.

Good translation of the more common languages would be nice, but simple translations, even - of a variety of languages would be really useful.

Pre-emptive strike

Since it's become "hip" to bash Google these days and support either MSN's search technology or Yahoo, I'm making a pre-emptive strike for the IT fashionistas:

"Duh!!! The best machine translator in the world already exists and there can be no improving upon it! Babblefish (thank you Altavista) has been doing this for well nigh a decade. All you Johnny-come-latelys are probably going to rave on with fanboy adoration of Google (the company that can do no wrong)!!! To top it all off, you lot apparently know nothing about Microsoft's language transtlation project which is slated to be deployed as part of Longhorny in 2010. Online language translation from Google will fail because Microsoft will have it built into the OS itself. Why send your document online for translation when the OS itself will not only translate it, but it will correct the grammar, punctuation and generate a WMA file in one of ten thousand gorgeously rendered synthetic voices. Google has lost. Google as been trolled. Google will have a nice day".

We now return you to your regularly scheduled pos[tt]en.

Old news...

There is already a tranzilator [gizoogle.com]

T.Q.

The system has been trained using the United Nations Documents as a corpus.

Seems one could devise a TQ (tranlsation quotient) measuring the effectiveness of machine (or human) translators. Take any standard reading-comprehension test, a send its text material through the translator, and back ...and then compare the scores of subjects taking the resulting test vs. those taking the original.

(Before such translators make their way into, say, diplomatic circles, I'd sure hope there's some objective demonstration of near-infallibility...)

oh no!

I don't ever expect such translation to work perfectly, but taking existing phrases should lead to useful first drafts.

This will mean one less possible career for me, and fewer babelfish induced laugther moments.

As a fluently bilingual person, I often recognize expressions that were translated in Canadian government documents. "Anglicisme" is the word the french have for it.

There's subtlety to languages we may forever lose. Take for example:

"Je donne ma langue au chat" - "I give up (answering a riddle) instead of the more picturesque "I give my language to the cat". Well, that should be tongue, but hey, it's just babelfish!

"Bullshit" won't produce "merde de taureau". That is a strange expression you anglos have, don't you realize?

"Il pleut comme vache qui pisse" will give us "it's pouring cats and dogs" rather than "it's pouring like cows' a'pissin". The french also have never heard of cats and dogs falling from the sky.

While an improved Babelfish may improve our mutual comprehension, please pause for a moment to consider all the linguistic hilarity we'll forever lose.

Re:oh no!

While an improved Babelfish may improve our mutual comprehension, please pause for a moment to consider all the linguistic hilarity we'll forever lose.

Yeah, like me going to work for Bull [bull.com] in 1997, and searching for "comment dit-on, le, fuck, le chose sur lequel on tappe, thingy qui connecte a l'ordinateur, ah yeah, le clavier". French Bull dude: "ah, le keyboard."

Hilarity indeed.

Time to move the AI bar

First, this is outstanding; Google, unsatisfied with traditional machine translation techniques, pioneers their own design. I'm certain their advertisers will be pleased to have their adds auto-translated to whatever language is necessary.

Second, I think we'll witness a case of having the AI ante upped once again when another traditional AI challenge is met. Wikipedia puts this best; When viewed with a moderate dose of cynicism, AI can be viewed as 'the set of computer science problems without good solutions at this point.' Once a sub-discipline results in useful work, it is carved out of artificial intelligence and given its own name.

Lovely translation source...

So when you go to translate.google.com and translate something, the result will be legal-eze in the resulting languages.

Spanish: "Que pasa?"
English translation: "With regards to the current situation, how is the day progressing?"

Re:Lovely translation source...

No, it actually translates "que pasa" into "We hereby condemn these actions taken by the Israeli government."

DVD's subtitle tracks

DVD subtitle tracks would be another good addition to help pick up slang too (most have an english track along with a couple others depending on the region)... all time-synced and easy to match up...

(I'm guessing that it'd fall under fair use and google wouldn't have to struggle to get the movie studios approval, (even though such tech would benefit the studios too))

Starting Wars !

In 'Hitchhiker's Guide to the Galaxy' (the 'trilogy' of books, not the recent movie), it's mentioned that the babelfish has effectively started many, many wars. The reasons seem to be that any being can be rude to any other being without a serious set of translations that explain exactly what the rude terms mean and how they should be regarded.

I'm highly concerned for this warmongering that Google has undertaken.

Reference Here: http://www.bbc.co.uk/cult/hitchhikers/guide/belgiu m.shtml [bbc.co.uk]

Picture this: I write a blog entry with either bad punctuation or erroneous content. Under the old system (pre-Goolge translation), I would receive several flames about my idiocy. With Google translations:

* People around the world will be confused and angered about my punctuation;
* Vastly larger numbers of people will complain about my erroneous content;
* Other people will step up to my defense and a massive flame war will ensue;
* Idiots eveywhere (who speak other languages) will echo my idiocy by believing the erroneous content I posted;
* The signal to noise ratio of the net will rise markedly;
* I will still be unsure of whether to count on my fingers starting with my thumb or forefinger depending on which European country I'm in.

I believe this pro-war, anti-peace, conflict-ridden idea of making everyone THINK they understand each other is ripe for critism. God made everyone else speak funny, I think it should stay that way! Only right thinking people speak my language anyway, and everyone else should just shut up and sit down!

(WARNING: above post contains carcinogenic levels of sarcasm, fasciousness, satire, irony, and adjectives. Please unplug brainstem and wipe with a clean, damp cloth before continuing.)

yeah, but can it translate this?

Wenn ist das Nunstruck git und Slotermeyer? Ja!... Beiherhund das Oder die Flipperwaldt gersput. be careful! If you translate this you may end up dead.....

words don't really have meanings

Some people here seem to have a false picture of how language works. Individual words do not have meanings. Not to a human interpreter anyway. Sentences used in actual contexts have meanings (unless a single word is uttered as an elliptical sentence). The "meanings" of words, as found in dictionaries, are simply abstractions from occasions of use. The idea that individual words have meanings hasn't been current in philosophy or linguistics for about 50 years. Also, the idea of St. Augustine that children learn the meaning of words by associating sounds that they hear with particular objects that they observe is now also considered rather dubious.

How will it translate ambiguous headlines?

I smiled when I read this recent headline:

Clinton tours devastated Bandeh Aceh.

Of course, I knew what the writer really meant. But the Bable Fish translation into French produces exactly the meaning which I first parsed when reading that headline.

Les excursions de Clinton ont dévasté Bandeh Aceh.

If machine translation become more common, perhaps English writers will have to be a little more careful.