Can Translators Learn to Love MT?

Created on www.canva.com

First things first: we are not experts on machine translation (MT) by any stretch of the imagination, but we certainly know that it's an incredibly polarizing topic in our industry. It is, perhaps, the most polarizing issue we've encountered in the last 10 years--in addition to how much we should charge for our services, of course. Now, our lovely colleague Jost Zetzsche is, without a doubt, the expert in the subject, and he told us about Lilt. Jost wrote about Lilt very eloquently here. Since the recommendation came from someone we trust, we went and checked it out. For the record: MT is integrated into many of the translation tools we use (Smartling, Wordfast, etc.), and while most of the MT-generated suggestions for translations are still quite giggle-inducing, MT has been getting better. As we've mentioned: we don't know that much about MT, but we have, chosen to view it as a tool to help translators become more efficient rather than a tool that will replace us. Others view this very differently. Allow us to elaborate on Lilt. And you probably know this, but let us emphasize that we have no financial or business connection to Lilt. We just think it’s a cool tool.

Courtesy of Lilt. Great summary indeed.

We took Lilt for a spin, and here are our findings: the program is very promising. Basically, it's a super-easy browser-based interface, there's nothing to download (just sign-in with a Google account or e-mail address). It's technically a statistical machine translation tool (SMT).  It was developed by the smart folks at Stanford Natural Language Processing Group, and it learns as you translate. We tried it for English->Spanish, and while it created some hilarious results, some were also very good. Lilt is currently available for EN<->ES, EN->FR and EN->DE). It's very nice to know that this system was created by a bunch of brilliant programmers at top-notch universities who are a lot smarter than we are. The whole idea is to make translators' work better and smarter through the use of technology. We know this is a big shift, but think about it: we work smarter and better because of computers in general, right? (Trite, but true.) They've allowed the industry to flourish, in part because we can research better and don't have to type up our work on a typewriter anymore (not that we are old enough to remember this). Perhaps this is the second revolution for translators--courtesy of computers, software, and computational linguists. Speaking of computational linguists, we had the chance to interview Spence Green, one of the founders of Lilt. He holds a PhD in computer science from Stanford (yes, smart guy; and nice, too). So here's a lot of information, straight from him, because he can say it so much better than we can. We sure do think we are ready for the next era of translators learning to embrace MT. What do you think?

Courtesy of Spence Green

Translation Times: What would you say to translators who are afraid of MT? And where does Lilt come in?

Spence Green: Machine translation is a tool. Specifically, it is generalized translation memory. Whereas a TM can only produce an output for an input that matches X% of some previous input, a sufficiently large MT system---where all of the major commercial systems today are sufficiently large---can produce an output for any input. The intellectual heritage of TM traces to Martin Kay's 1980 position paper on man/machine approaches to translation. Of course, Martin was and is an MT researcher. So translators should see MT as an evolution of a tool that they already use and ostensibly derive value from.

Lilt is an interactive MT system (https://en.wikipedia.org/wiki/Interactive_machine_translation). Interactive MT is an old idea that dates to at least the 1960s, but it has never been widely available for translators. Lilt is based on a research system we built at Stanford called Predictive Translation Memory (PTM). PTM began as a human-computer interaction project to understand translator behavior in the presence of machine suggestions, be they from TM or MT. It evolved into an interactive MT system. About 100 professional translators from Proz and elsewhere participated in the three research studies during 2012-2014.

In my view, the issues that seem to upset translators---the MT post-editing experience, rate discounts, often unrealized promises of productivity improvements by vendors---have more to do with the peculiarities of the industry than with the technology itself. 

TT: Why did you create Lilt?

SG: After undergrad I moved to Abu Dhabi. This was in 2005, and I was 24 and restless. I wanted to learn a non-European language, and it seemed to me that the Arab world would be very important during my lifetime.

Two significant things happened during that time. First, after about a year, I found that most of my friends were Arabs who didn't speak English well. Through their eyes I saw a different world. You often get paid less. There aren't as many books available. Wikipedia and Google search stink. Information access began to matter to me.

Second, Google Translate was released. I thought that this technology could solve the information access problem at a scale that human services never could. I wanted to learn how it worked. But I soon realized that my undergraduate training was insufficient. So I applied to Stanford and started graduate school in 2008.

In 2010 Ed Bice of meedan.org gave a talk at Johns Hopkins, where I was spending the summer doing research. He runs a small translation non-profit in San Francisco that uses MT from IBM. I started to visit his office regularly. I began to understand the differences between the assimilation and dissemination use cases of MT. Google Translate solves the former, but the latter---which is translation with a quality threshold---seemed like an underexplored area consistent with my desire to improve information access.

Courtesy of Lilt

I spent the summer of 2011 at Google working on Translate. I learned how to build huge translation systems. And I learned how many people send translation feedback to Google, and what can be done with that feedback. There was real work to do that was both intellectually compelling and high-impact.

At Google I worked with John DeNero, who worked for Franz Och. John and Franz both left Google in 2014. I graduated at the end of 2014. John is a co-founder of Lilt, and Franz is chief advisor.

TT: Wow, that’s quite a pedigree. Now, what’s your current revenue model?

SG: Lilt is in a free trial period. We will add a professional version in 2016. We charge for our translation API, for which we currently have one launch customer.

TT: Sorry for this question, but: what’s the catch? If Lilt is free now, how will you make money? 

SG: Yes, the professional version will be released later this year. But there will always be a free version or a free trial.

TT: Can you tell us when exactly other languages will be added?

SG: Fr-En and De-En will be added next week. En-Pt will follow in February. The Danes and the Dutch seem to be pretty enthusiastic, so we'll probably add those language pairs in February.

TT: Can you tell us about yourself, your background, and your core team? Who are you? What’s your favorite author? Your favorite musician? 

SG: I'm a Southerner, but I've moved around a lot since high school, so my accent doesn't betray my nationality anymore. I grew up in Atlanta, went to U.Va, and then to grad school at Stanford. I speak Arabic; I run; and, I am a scuba diver. My favorite book is The Name of the Rose by Umberto Eco. If I were stuck on an island forever, I suppose that I wouldn't tire of listening to The Joshua Tree by U2.

John DeNero, my co-founder, completed graduate school at Berkeley, then spent four years at Google on Translate. Now he is a CS professor at Berkeley. He teaches the largest class at the university. He's a busy guy.

Joern Wuebker and Sasa Hasan, the other two founding members of our company, completed graduate school at RWTH Aachen under Hermann Ney. They're great. Hermann produced Franz Och, who then worked with Philipp Koehn, then a grad student at USC. Franz and Philipp went on to produce Google Translate and Moses, respectively.   My brother Richmond is in charge of marketing. It's great fun to work with a sibling, as you must know! Chris Manning and Jeff Heer, my two academic advisors, are also on our advisory board.

TT: Did you raise any venture capital or how is your company financed?

SG:Yes, we raised a VC round. We are backed by XSeed Capital.

TT: We know you don’t have a crystal ball, but do you think translators will ever be replaced by MT? Is Ray Kurzweil right? What’s your hunch?

SG: This is a philosophical question that Nietzsche, Quinn, Jakobsen, and others have considered. I tend to think that their treatments were more sophisticated than those that you read in the media. I wrote about this in the introduction to my dissertation (sect.1.2; http://www.spencegreen.com/pubs/dissertation.pdf).

Short summary: yes, for text that can be memorized or otherwise easily routinized (imagine a large TM of all of the translations ever produced in the world...). Not anytime soon for any sort of translation that requires world knowledge. That problem is AI-complete and would signal the advent of artificial intelligence. 

  

TT: What’s the most underrated website on the internet, in your opinion? 

SG: Wikipedia. It is one of our greatest intellectual achievements as a species. 

Many thanks to Spence for answering all of our questions. We've barely scratched the surface with Lilt, but will continue playing around with it. It seems like a great tool that can potentially really make every translator's life easier, and as scary as MT can be, we encourage colleagues to keep an open mind and view the tool as a benefit. And no, we are not about to work as post-MT editors anytime soon (we actually don't know anyone who does). What do you think, dear colleagues? Do you have any questions for Spence? We bet we can twist his arm to answer them here. 

Related Posts

Enjoyed reading? Share this

• Technorati
• Delicious
• Digg
• Facebook
• Twitter