The problem with the world wide web is that it isn’t exactly worldwide. For many users, their internet doesn’t really extend beyond the borders of their particular country, and in most cases, this is simply a problem of language. Websites from across many boundaries just aren’t written in words they can understand.
EBay is just one of many companies aiming to break down this language barrier, but it has an especially strong reason to do so. The more people that eBay sellers can sell to, the better chance of making a sale. This ultimately means more money for eBay. To help widen its markets as much as possible, the company is now testing an automated translation tool on its product listings. In the process, it hopes to improve the art and science of online translation in a way that few other sites can. EBay engineers believe they have an especially good shot at developing effective translation methods using a key eBay-centric data point: whether something sells or not.
Hassan Sawaf became eBay’s director of machine translation a little less than a year ago, after a long career researching and developing ways to make computers better at understanding human language. The job attracted him, he says, because so much work in the field takes place in the abstract. Translation algorithms are trained against reference texts and scored on how close the computer version gets to the human original. At eBay, he says, he gets to see how machine translations fare not compared to how well they mimic human text or speech, but by how humans respond while interacting and transacting in the real world.
“In our case, we care about what is giving you the signal as a user to basically make the acquisition, buy the product,” Sawaf says. “We are not optimizing toward automated scores, which are almost artificial. We are optimizing toward the user behavior, the user experience, the user action.”
Sawaf’s first big experiment is a translation tool now being beta-tested on eBay Russia. Basically, it displays the title of each listing in Russian by default. That may sound like a modest beginning, but consider a recent search for shoes alone on the site returns more than 3.6 million results — results that are constantly changing as new listings are added and old ones disappear. The tool translates titles into Russian from multiple languages, including English, Spanish, German, and Chinese.