Translation Guy Blog
Machine translation has improved by leaps and bounds. What was once considered machine-produced gibberish is increasingly giving human translators a run for their money, particularly for predictable texts like weather reports. While machine translation (MT) is also more economical than human translation, it's not a true alternative yet. In most cases, machine translation can't be used as is. And that's where the expertise of machine translation posteditors comes in. Machine translation posteditors are the human editors that work to improve the output of machine translation. They combine the MT output with their linguistic expertise to provide a better reading experience to human audiences. Besides the cost savings, it is estimated that machine translation plus postediting is 40% more efficient than human translation alone. But what exactly do machine translation posteditors do, and how do they do it?Read more
eBay has started working on a new approach to machine translation (MT) that could represent a breakthrough in machine translation quality. A team lead by artificial intelligence guru Hassan Sawaf, renowned for his work on hybrid machine translation systems, which combine different translation engines for better translation results. Sawaf calls what he’s trying to do for eBay “context translation,” since the new MT intelligence looks at a lot more than just a sentence when composing a translation. Russian is now complete and has provided a template for roll-out of additional languages that eBay hopes will create a multilingual market for eBay sellers regardless of their preferred language.
Leena Rao at TechCrunch describes the challenges eBay faces in emerging markets like Russia. “eBay is trying to curate inventory from a global base of sellers and surface this to buyers in emerging eBay markets based on what ships to them in their respective countries. A Russian user can go to the localized version of eBay and see all products that are listed in Russian. When they are inputting search terms in Russian, this engine will produce search results of listings that match the query in Russian. But the Russian user’s query will not be able to see posts that match their query that were written by sellers listing in English. In order to access English listings, which do represent a considerable number of the listings on eBay’s platform, Sawaf explains, the Russian user would have to input the query on eBay in English.
“Machine translation normalizes this,” says Sawaf.
With over 100 million products listed at any one time on eBay, pricing changes by the second, and lots of old stuff sold by amateurs and shipped around the world, there is a lot of back-and-forth between buyer and seller, which creates a lot of room for misunderstanding and missed opportunities. eBay thinks they can do better with their own in-house context translation solution.
This because the big database-driven statistical engines draw across the entire internet for their translation matches. Sure, using the sum of all human knowledge, (which is what language is, if you think about it), seems like a great way to find the perfect phrase in translation, but it’s not. This because the same phrase can mean different things in different contexts or domains, as they are usually called in the business. (I have to confess that I am not certain if Sawaf’s “context” is about domains or something else.) I always use the lame example of “The server is down” when I talk to my clients about this. Depending on context, that could mean that a computer has failed, or that a waiter has taken a spill. (I’m looking for a better example, if you got one, please post it.) The eBay user experience provides a unique domain, or really many domains depending on product category.With all those millions of transactions, eBay has an enormous amount of data specific to that user experience, which can be used by the machine translation AI to provide more accurate translations specific to the language of eBayers, no matter what their native language. Spanish and Portuguese are next. eBay press release here. Readers, if you want more on this, comment below, and I will reach out to Sawaf.Read more
Real-time speech translation will revolutionize how we communicate, according to Jonathan Luff at Wired magazine. “Having thought about the issue a bit, and having witnessed some extraordinary advances in artificial intelligence, machine learning, wearable technology, and real-time data analysis, I have come to believe that the next big leap forward is not represented by Google Glass, … but by ‘Google Ears’.” Thanks to “Moore’s Law”, Luff believes that accurate, real-time translation of spoken language is just around the corner. “Just think what an impact this will have. You will be able speak to, understand, learn from, and do business with anyone, in any place, at any time.” Sound familiar? I mean, I’ve been listening to this ill-kept promise for 20 years. And despite all the hyperbole, machine translation and voice recognition have been getting better and better. But that Babel Fish paradigm, that instantaneous, Star Trek-like communication so useful to Captain Kirk’s xenophilic predilections remains out of technological reach So maybe Luff knows something I don’t. After all, he’s the one that got the editorial in Wired magazine. They won’t even send me a rate card, for crying out loud. Luff, a former diplomat, used to be British PM David Cameron’s advisor on digital strategy. But he recently quit that job to lobby in the UK for Wonga, a payday lender specializing in short-term online loans. With an APR of 4214%, it is loansharking made legal, and certainly qualifies Luff as a shill for digital pocket picking. So first as a politician and then as a lobbyist, Luff has certainly proved adept at vacuuming the pockets of his clients, but how this qualifies him to comment on the bright new future of machine translation perplexes me. He has “thought about it a bit” in light of a project to translate more books into Arabic, possibly. But for Wired editors, it’s not the messenger but the message. Instantaneous translation – that little technological G spot, the chance to speak any language virtually without actually having to know what you’re talking about. The nitty-gritty of actual translated speech, the misunderstandings, the headaches, the pure discomfort of the careful thought required for cross-cultural communication solved by the shiny new translation instrument is not considered at all. That kind of stuff is too real world. Better to look at things virtually through Google-tinted glasses. It’s irritating that I am participating in this bull ship process, too. First you have the bright promise of the future. Then you have the curmudgeon saying it will never work. So keep moving. No news here. The only real, and irritating, consequence is the way this Star Trek communicator fetish complicates my own sales pitch. The hucksters talk about translation technology as if it’s the best thing since sliced bread, leaving us bakers to explain that the flavor of the bread is not in the slicing.Read more
The biggest problem with machine translation on the Internet is not that it’s inaccurate or can’t translate poetry or Shakespeare. The problem is that it’s pull technology rather than a push, meaning you have to fetch your translation rather than have it come bouncing to you.
Rick Rashid, Chief Research Officer at Microsoft Research, demonstrated the latest breakthrough in speech recognition and machine translation a few weeks ago by giving a speech in Taipei using real-time computer-generated Chinese audio translation. Rashid kicked off in English, providing a great summary of the history of machine translation and voice recognition. It was a good overview of the 60-year effort to build computer systems that can understand what a person says when they talk, and to translate what they said. Way back when, voice recognition started off with simple pattern matching of voice prints. Because each speaker’s voice was so different, it was hard to recognize speech that deviated even slightly from the pattern. Later, scientists programmed statistical speech models constructed from the recorded voices of many speakers. The software used to integrate these voices is known as hidden Markov modeling and was the breakthrough needed to get the ball rolling. In the last 10 years, better software and faster computers have led to more practical uses. Now it seems as if machines do most of the talking on the phone, but their capabilities are still quite limited, as we all have frustratingly experienced. Even the most robust systems are still reporting error rates of around 25% when handling general speech, according to Rashid. Machines do a lot better when they’ve been trained for an individual voice. A few posts ago I blogged about my own experience writing this blog by dictation. Untrained tools remain error-prone. Researchers at Microsoft Research and the University of Toronto have applied a new technique, called Deep-Neural-Network Speech Recognition, which is patterned after human brain behavior. Results were about 30% better. According to Rashid, “This means that rather than having one word in four or five incorrect, now the error rate is one word in seven or eight. While still far from perfect, this is the most dramatic change in accuracy since the introduction of hidden Markov modeling in 1979 and, as we add more data to the training, we believe that we will get even better results.” Note that this increase is without the speech adaptation required to improve earlier systems such as the one I rely on. At 6:41 into the video, Rashid begins to use the tool, which has been modified to match his voice. The tool transcribes his voice, translates to Chinese, and then reads it out loud in Chinese programmed to match Rashid’s voice in English. The affect is uncanny, and the Chinese-speaking audience received the translation with enthusiastic applause at the apparently successful translation of each simple line, translated slowly and consecutively. It looked really impressive. Rashid blogs, “Of course, there are still likely to be errors in both the English text and the translation into Chinese, and the results can sometimes be humorous. Still, the technology has developed to be quite useful. “Most significantly, we have attained an important goal by enabling an English speaker like me to present in Chinese in his or her own voice, which is what I demonstrated in China.” I have no way of telling if the translation is any good, so I encourage our Chinese-speaking readers to listen in and report. But quality might not be that important. This kind of tool doesn’t have to be good, just good enough.Read more
On the treadmill with voice recognition and machine translation. NTT Docomo, Japan’s leading cell phone carrier, has unveiled its real-time speech-to-speech translation service, reports Geek.com. As a supplier of telephone interpreters, I’ve been waiting for this for a long time, since I used to imagine it would put us all out of work. But I’ll hold off on searching Craigslist for a new job just yet. Fact is these tools are just window dressing, press release fodder for the uninitiated, because they just don’t work that well. I haven’t tested the DoCoMo tool, but I’m sure it’s the same old tech BS. How do I know? Because I’m using the best speech recognition tool available to compose this post. That’s right, no more typing for me. It’s all voice recognition for TranslationGuy. Tendinitis has ended my typing days. I’m using Dragon Naturally Speaking 12, the latest and greatest voice recognition tool from Nuance. It is an amazingly powerful technology, a real game-changer. I’m writing twice as fast, with half the pain, after only a few weeks of intensive frustration. Thinking before speaking does not come naturally to me. Nuance technical support tells me that the tool will never be able to transcribe the name of 1-800-Translate because of the unfortunate use of hyphens in our brand. And to Dragon’s tin ear,“Ken” remains indistinguishable from “10” or “can,” no matter how many times I program the tool. Unless I change brand and name (any suggestions welcome) a careful post-edit is always required. Which reveals the big dirty secret behind Apple’s OS, and it’s not that the maps are stupid, which I guess is not really a secret. The secret is that Siri sucks. Apple’s voice-recognition tool is powered by the same Nuance engine I am struggling to train right now. Having spent hours training it to my voice, and edited at my hand, it is an amazing tool, but when applied to real-time communication the error rate is too high. Siri may be geek-sexy, but her output does not put out. This is great for my business, because it means our telephone interpreters are going to be around for a long time. We still turn the wheels of real-time translation, just as Fred Flintstone’s saber-tooth squirrels on their treadmills kept Stone-Age Bedrock rolling. Yabba dabba doo! Without training first and post-edit last, voice recognition and machine translation are just useful-ish. Combining two halves of two incomplete processes just doubles your troubles. But once we get some saber–toothed linguists into the automation dust-up, wonders can be achieved. If I could, I’d make everyone on my team use them all the time. The training would be killer, but the payoff is we would all make a lot more money, at least until the other guys figured it out. Mastering these automation processes requires patience, persistence and practice (one of my Dad’s favorite lines, although I think he was mostly referring to sex whenever he said it). To make good use of these tools requires planning and changes to old habits of work and thought. The only thing more painful is carpal tunnel syndrome, or business lost to tech-savvy competitors, which is even more painful to my wallet. Ouch!Read more
Spanish is the only good news in the news business. Spanish news consumers are growing in the US, while shrinking English-language readership consigns a once-mighty industry to a mere lining for Wall Street’s birdcage. So dying newspapers look to build new audiences among Hispanic readership by publishing in Spanish. Lots of all-American news brands now have dedicated Spanish-language sections. The Hartford Courant found a way to both save money and avoid readership decline, and that’s by translating the paper’s home page into Spanish using Google Translate. “The limitations of this approach are immediately apparent to Spanish-speakers,” wrote Andrew Beaujon of the Poynter Institute. There was a public outcry over the machine-made service to the Hispanic community. So the Courant published a disclaimer on the lousy machine-translated Spanish, which stated, “Some of the translations of the English headlines and articles don’t always translate accurately word-for-word into Spanish.” “In an attempt to improve the translation service, “Google has included a wiki/crowdsource feature that allows bi-lingual users to write better English translations for each article,” the Courant wrote. “Simply hover over a story with your cursor, enter the translation and help write a better English to Spanish translation.” Just the web experience newsreaders are looking for when they go to check out local events. Fixing Google Translate. The response rate must be almost zero. Bessy Reyna of ctlatinonews.com thinks this is a problem. “I still think that if the Courant wants to truly offer a product that provides information to their Spanish-speaking readership about their community, they could at least hire someone to translate the translator. The guessing game (trying to figure out what the Spanish translation means) is painful and time consuming. Google recommends that each reader “improve” the translation using wiki / crowdsourcing. Thanks, but no thanks. It would take hours to fix the many problems found in each piece.” My post MT editors say the same thing about machine translation. It’s hard to follow, and harder to fix. Looks like the Courant has provided the kind of translation that even readers won’t touch. So is a horrible translation better than none at all? Hispanic media consumers are heartily sick of horrible translation and the Courant en Google Translate Español has to be among the least likely sources a Spanish-language reader would turn to online. I should disclose that I was formerly employed by the Hartford Courant as a newspaper delivery boy. Actually, I was a subcontractor to my big brother, and delivered half his route for him on Sundays, for which he paid me $0.25 a week. If that doesn’t seem like much money even for 1966, you’re right. I think he was cheating me From that perspective, I think the Courant is cheating Spanish readers, too. But good or bad, utility is determined by usage on the Web. If people use it, then it is useful, in absence of a better solution. If the market doesn’t work for translation, it may be better to do nothing at all. It would be interesting to talk to the Courant people and get their perspective.Read more
Back in the day at CNN, you could be sure that field producer in Moscow spoke Russian. For old-school foreign correspondents, it took a second language to get the story. In the new media, the social media, that kind of local knowledge is both no longer sufficient and no longer required for breaking news argues CNN’s old Moscow field producer, Mike Sefanov. Mainstream news war-horses have been left at the starting gate when it comes to breaking news. Wise old pipe-smoking bureau chiefs have now been put out to pasture. “Today’s online translation tools make it possible for anyone with an Internet connection to read texts in foreign languages and translate entire web pages. The addition and proliferation of social media has made our world even smaller by allowing us to correspond with one another on a personal level across barriers previously penetrable only with the aid of interpreters,” says Sefanov. For journalists like him, “the convergence of these technologies means that we can travel from story to story without stepping away from the computer, interview locals without hiring a fixer, and gather content from the epicenter of an event.” Sefanof provides a blow-by-blow example on a breaking Syrian story using Google Translate. “I came across the following tweet from local activist @HamaEcho: “Burning the political security branch & municipality http://www.youtube.com/watch?v=rnkDcirFDf4&feature=plcp … & police stationhttp://www.youtube.com/watch?v=LhEw0LwcJXU&feature=plcp … of Hajar Aswad, #Damascus — Sami al-Hamwi (@HamaEcho) July 18, 2012 “The tweet identified the video as taking place in Hajar Aswad, Damascus. Google Maps shows us that Hajar Aswad is a neighborhood in Damascus proper. But the title of the first YouTube video linked to reads, 'حرق مبنى المجلس المحلي لمدينة الحجر الاسود 18-7-2012.' Because I do not read Arabic, I used Google Translate, to obtain the following: “Burning building of the local council of the city of the Black Stone 18/07/2012”. This translation confirmed what @HamaEcho described, except for the reference to “the city of the Black Stone” But by translating “city of black stone” using Google Translate, and then pasting the Arabic translation into a Google Map search, Hajar Aswad was confirmed. Saefanov has souped up his Chrome browser to use the translate extension to follow foreign languages conversations more easily. “This is crucial: in most cases, those “reporting” on the ground frequently do so in their native language,” he says. For anyone doing research outside the English ghetto, setting up Chrome to translate non-English pages automatically creates a whole new surfing experience. Sefanov iseaven using the same technology for mini language lessons. “By taking a word and placing it into Google Translate and then pressing the “listen” button, it is even possible to hear the way something in a foreign language sounds. This becomes useful in instances where activists might be naming the location of what they are filming within the video itself – a keen ear can pick up the place name to help in the process of verification.” Sounds like a stretch to me, Mark, but maybe you’ve got the ear for it. For textual translation it can be invaluable for regular non-English searches. I agree with Sefanov that none of this is a worthy substitute for the first-hand knowledge of a local ways and insider expertise, but tight deadlines and tight budgets are changing the way the news is made. Without journalistic wing-tipped shoes on the ground, MT tools have been used to replace the irreplaceable. "In any case its very important to rely on local knowledge to make sure that viewers understand the context of what they see on YouTube or Twitter." Mike provides a really invaluable description of how a professional advances his career through the use of the technology, and I thought some of his MT work-arounds were very clever. Anybody else have some success stories on how they use automated translation tools for research in other languages?Read more
The largest translation effort in the world services 200 million users a month, translating the textual equivalent of 1 million books a day. Can you guess who that might be? With numbers like that, you know its gotta be Google. Franz Och, Distinguished Research Scientist, (his title, not mine) at Google Translate marked six years of progress in a recent post at Google Translate Blog. In just one day, Google translates as much as all the professional translators in the world translate in a year, says Google. A million books multiplied by 100,000 words per book is 10 billion words. Only Google has the rack space to run this kind of volume, and the technology to do it efficiently. When they started out, the distinguished Och reports that it took 40 hours and 1000 machines to translate 1000 sentences, which for those of you lacking a technical background means that it sucked. “So we focused on speed, and a year later our system could translate a sentence in under a second, and with better quality. In early 2006, we rolled out our first languages: Chinese, then Arabic.” Now Google offers 62 languages total, and has a pretty good kit for the less commonly translated languages, so many more to come. While those human translators billed billions for their billions of words, Google just gives the translations away, although it now costs money to license the app, at something like a penny or two per page, my best guess about .03 % the cost of human translation (Note: that’s 3/100 of a percent, not 3%) And a tip of the hat (from me, not the distinguished Och), to all of you who spend your lives translating words for a few pennies each. Thanks to your for providing all the good, paid translation that Google then scoops up in order to turn into bad, free translation. As the distinguished Och notes, “we believe that as machine translation encourages people to speak their own languages more and carry on more global conversations, translation experts will be more crucial than ever.” Note that he said “translation experts” not “translators.” But dollars and cents really don’t give a sense of what a transformative technology machine translation has become thanks to Google’s reach. Google Translate reports that 92% of queries originate outside the Unites States, and that use of Google Translate on mobiles is increasing at four times the rate of desk-bound systems. So Google Translate will become even more ubiquitous in the future than it is now. Google has taken their current statistical approach as far as it can go, say I. For years statistical machine translation guys have been asking for more data. But Google Translate reached critical mass a long time ago, and even vast new stores of data have had only the most incremental impact on the quality of their machine translation. Alexis Madrigal argues in Atlantic that Google is going to have to come up with some new tricks to make better use of the data they already have. “Google (or any other translation software) will have to start understanding (in some way) the semantic content of the words it is arranging.” Which strikes me as a kind a knuckle-headed comments that at least demonstrates that his blogging heart is in the right place. Ilia Kaufman of NoBable was developing AI algorithms’ to identify domain (subject area) in text as a means to refine machine translation output. So that when the computer translates “the server is down” it will be translated one way for an IT text, i.e., “the computer has failed,” and another way for hospitality industry, as in, “the waiter is injured.” That’s where I would put my money. But NoBable went out of business last week. You can see Google Translate in action against Bing and BabelFish at my Free Translation Challenge. I thought this page would be a big hit, but it’s not as easy to give away free translation as it used to be. Thanks to Google Translate.Read more
Siri, the app that talks back, is the big hit on the new iPhone 4s. In my last post, we looked under the covers at the unnaturally intimate relationships we furtively pursue with our cell phones. I'm talking about more than just an extra "s" in front of "texting." Our little hand-held fetishes have become the sacred mirrors from which all is revealed, as our communication with our smart phones evolves/devolves to a conversation with a device rather than the person at the end of the line. And now that apps like Siri are learning to say just the right thing, its just going to get better and better, or worse and worse. Youtube is awash with iPhone geeks eager to share their interactions with their oh-so-clever cell phones. Talk about stupid pet tricks. Thousands have put Siri through a Turing torture test to seperate the men from the machines, or at least to tests a machine’s ability to mimic human behavior. Yawn. Been there, done that. Even a casual viewer of Jeopardy knows that IBMs artificial intelligence, Watson, can already whup humans at facts and figures. How hard can being human be? We are only a data set away from a machine that can represent as more human than a human. Everything will be fine, so long as you've got Siri in your pocket. Who needs friends when an app has got your back? But what about that pal of yours at the other end of the line? How do you know it's not his Siri? How does it matter, really? Maybe that's who you should be talking to anyway. Rick Bookstaber figures its going to happen any day now. “One reason…, is that we are meeting the computers half way. The more we become twittering, texting beings, the easier it is for a computer to mimic us, because we are stripped of much of our human context and behave more like computers. “The second reason is now readily apparent with the unfurling of the Apple iPhone4S and Siri, the digital assistant. With the iPhone users accessing Siri to find restaurants, make appointments, and ask trivia-level questions (and with more areas of interaction added down the road), Apple's servers are going to amass the queries of millions of people many times every day. And as Google has shown with Google Translate, if a computer has enough raw material, it can pretty much figure this sort of thing out." So when we can’t tell the human from the ghost in the machine, it will be just like online poker, where poker ‘bots are lurking to clean your clock. Or skim just enough to avoid detection by anti-poker bots. And that's just the start. Artificial intelligence follows the money, and online poker is chump change compared to the revenue these apps will someday produce for their masters. The voices in the machines will help you so much that you will be helpless without them. And soon enough, the machine pretending to be a human will be a better at being human than you. They might even be better at being you than you, but just the way you like it, because they’ll learn that too, as natural language algorithms pour over your natural language as you speak. What will become of us?Read more