Intel, the world leader in silicon innovation, develops technologies, products and initiatives to continually advance how people work and live. Founded in 1968 to build semiconductor memory products, Intel introduced the world’s first microprocessor in 1971. Today, Intel has over 82,000 employees globally. Based in Hillsboro, Oregon, Will Burgett is the product manager for Translation Innovation & Services in Marketing Solutions at Intel.
Thicke: Intel is truly an innovator in this space. How is it that Intel was already deploying machine translation (MT) while others were just talking about it?
Burgett: I actually started to look at MT options in 1997 when I was a manager in Intel’s International Product Development organization. But at that time, the better-performing, higher-quality MT systems were still on mainframes, which were extremely expensive with a negative return on investment (ROI). It was only much later, in 2007, through involvement with the Translation Automation User Society (TAUS) that we encountered promising and affordable statistical-based MT solutions.
Thicke: When did you first start using MT, and for what kind of content?
Burgett: We started using statistical machine translation (SMT) in 2007 and developed a pilot system for Latin American Spanish. We integrated that system into the Intel customer support website to deliver and publish raw translation (no human edit). The support knowledgebase was so huge that with just a human translation option, we had only budget enough to translate a moderate percentage of the English into the different target languages. Today, of the ten languages on the Intel customer support site, we deliver fully automatic useful translation in five of those languages: Latin American Spanish, Brazilian Portuguese, Russian, Simplified Chinese and Korean. Fully 97% of those languages are done by MT, while 3% is done by human translation on content for legal, safety, warranty and some new products to increase translation memory (TM). We also have the MT system integrated with our translation management system for MT plus post-edit.
Thicke: What were the gains of using MT? What kind of savings resulted?
Burgett: The most important gain was giving customers access to all the customer support information in their languages. Even if it isn’t always perfect translation, it gives them a more consistent experience so that they aren’t navigating and browsing first in their language and then back into English. We have a number of methods to measure customer satisfaction with machine translated content, and those satisfaction rates compare well with the baseline English measures. In some MT languages (such as Simplified Chinese) we even show higher levels of satisfaction than with the English. Bottom line, our customers are happy to have all the content in their language, and they are able to get their information and solve their problems.
The second benefit is the huge cost avoidance and cost savings. We’re able to output three times more translated content at half the budget.
And last but not least, a project cycle for customer support used to take ten business days, and now our project cycle for the MT languages is on an automated 24-hour cycle where any new content added to the customer support knowledgebase is automatically translated and published with no human intervention.
Thicke: But if this process is all automated, don’t you worry about random errors and really bad translations?
Burgett: Oh, yes, definitely; we still maintain good quality control and assurance practices. We conduct periodic inspections of our MT sites to look for problems, and we have our product support engineers in the different geographic regions looking for issues and submitting them to a bug-tracking system for investigation and fixing. We also have a mechanism in place for our customers who know some English, which easily lets them toggle back and forth between the translated and English pages.
Thicke: Software and products, especially global consumer products such as cell phones, may be localized into 10, 20 or 30 languages, but the online content that supports global clients is traditionally not localized into the same number of languages. That may mean whole markets that don’t have access to support content in their language. Why is that?
Burgett: It’s a universal problem for anybody who develops international products. Money! For some products, the margins may be so small, the competition fierce, or the volume of content too great to be able to afford human translation and still make a profit. That’s why translation automation is so critical. For example, the customer support site has 10,000 files that have to be translated, and without MT it would cost many millions of dollars more. It is also why organizations such as the TAUS Data Association (TDA) are so vital. Sharing TMs through TDA helps us train and customize MT engines, increase leveraging and reuse of translations, and improve translation quality by finding the right terminology.
Thicke: Can we as an industry determine precisely the percentage increase of sales that result in a market when support content becomes available in the local language?
Burgett: That’s certainly the Holy Grail for us GILT folks. Research by industry analysts points out that internet customers prefer buying from sites localized in their languages; developers favor buying from companies that deliver localized software tools and documentation; and most customers want their troubleshooting and support information in their language. Software companies can, and many do, measure the ROI on localizing their products for their target locales and the resulting increases in sales. Intel develops and delivers hardware, software, services and innovative solutions. In such a rich product ecosystem, with so many market and technology influences, it is a challenge to separate out the revenue impact of localization from all the other influences in that ecosystem. There are so many environmental variables that drive sales; it’s a challenge to isolate just the impact of support — and much more so the impact of localized support — from all the other variables and influences. But we are working on some ideas and techniques that may help us make those kinds of measurements in the near future.
Thicke: So now Intel is once again pushing ahead. Multilingual chat strikes me as the next great frontier for MT. What business drivers are behind Intel’s interest in automating multilingual chat?
Burgett: In general, chat is a contact medium that has seen tremendous growth in recent years. Customers seem to be more satisfied with chat than with other kinds of contact mediums, including phone and e-mail. Chat is also much less expensive, and a good agent can handle three or more customers at the same time. One of the biggest challenges using chat on a global scale is the ability to support different languages in a 24x7x365 mode and find the talented people who can speak all the key languages and have the technical skills. Of course, these are also very expensive resources. Multilingual chat can greatly enhance global flexibility and give us more options on where those resources are located geographically.
Thicke: What are the challenges you faced in developing an MT solution for chat?
Burgett: Of course there were a number of challenges, some technology-based, some human-based. Good integration between two very different applications is always a challenge, and such was the case with the MT system and the chat application. Neither was designed with the other in mind, so you have to come up with some creative solutions as well as some less desirable workarounds. In the chat application you have both the client and agent sides of the application, and both have to see the text in their native language as well as the translated language. In addition, we wanted to add a feature that allowed a reverse translation for real-time feedback on the quality of the translation. On the human side of the equation was the wide range of opinion on what quality is and what is good enough for the purpose of reasonable communication. We put a pretty good rating system in place to help our evaluators and quality assurance (QA) team score the quality of the real-time translation, but people still have widely varying opinions based on their likes and dislikes.
Thicke: Tell us about the pilot you ran.
Burgett: Our objective was to research the quality, performance and usability of using MT to deliver multilingual interaction capability for both customer support and sales. We developed three prototypes using different MT systems, but the same chat application. Our test languages were English < > Latin American Spanish and English < > Simplified Chinese. The team created eight different chat scenarios, for a total of 4,000 chats. Using a Likert system to evaluate the quality of the multilingual chats, we first did a baseline evaluation using our localization QA team. For the full evaluation, we recruited support agents from the Latin American Call Center and the People’s Republic of China. These agents played the roles of both customers and agents who scored the chats based on the Likert scale. Besides quality, we also collected data on performance and usability.
Thicke: What key metrics were you measuring? Were they the right metrics?
Burgett: Using the Likert scale for the quality scoring, we used 1 to 5, where a score of 1 meant that the translation was “Not Understandable” and 5 indicated the translation was “Understandable and Actionable, most text translated accurately.” Using some good data analysis techniques, we crunched the numbers and came up with a variety of different ways to look at the data. Lastly, we had all the participants fill out a survey and feedback form to get more detailed opinions, experiences and sentiment. I think for the most part these were all good metrics to use, although I think we could have used some automated measurements such as BLEU and NIST to get some contrasting measurement perspectives.
Thicke: In our work with rules-based engines, we find it pretty easy to make the leap to customer support content when the engines have already been customized on product terminology. With SMT for chat, were you able to use essentially the same engines you built for support? Or did you have to create new engines?
Burgett: Actually, we used a broad range of training methods and found that we got good results from all three of our prototypes.
Thicke: What were the results of your pilot?
Burgett: Our conclusion was that yes, indeed, we can get good results integrating MT into chat and that the performance and quality are good enough for customer support type environments. We also feel that the quality levels would work well in other environments, such as sales. But more needs to be tested specifically in that specific environment.
Thicke: Are you planning to also translate community content?
Burgett: We’re interested in multilingual collaboration and interaction opportunities of all kinds. Translating social media and dynamic content in real time is the big wave of the near future. Whether it’s Facebook, Twitter, communities or blogs, people want to communicate and to share opinions, ideas and experiences. They want to collaborate, debate and define, and they want to do it globally. User-generated content is infinite, valuable and ephemeral. Only translation automation in real time can tackle this tsunami of content. And far from putting human translators out of work, I think it’s going to create a whole new universe of business opportunities for them.
Thicke: What is the importance of community content to Intel?
Burgett: Social media, including communities, can build relevance in a message because it involves the audience and a voice of many over the voice of a few. We are building loyalty and brand sentiment through our communities by better understanding our customers’ needs that help us find the innovative products and solutions. Communities are “the big ear” to customer value, and we must be able to do that listening and communicating in their language.
Thicke: Do you intend to do any normalization of source text to transform abbreviations such as LOL and just plain idiosyncratic spelling so that an MT engine can better recognize the words?
Burgett: Ah, yes, this kind of unstructured, Wild West content has many perils for translation automation. Short answer: yes, we need to work on it using multiple techniques, including normalization.
Thicke: So, what’s the future for multilingual collaboration and real-time translation at Intel?
Burgett: Paul S. Otellini, our president and CEO, has said that “Intel has arguably an audacious vision: This decade we will create and extend computing technology to connect and enrich the lives of every person on earth.” And we will connect and enrich those lives in their languages. Translation automation will play a growing role in that vision.