Glaring errors can confuse customers, so companies are training computers to do better

When Etsy Inc. lists a “clutch” or “cross-body bag” for sale in its online marketplace, it’s usually listed as a bag. When machines translate those listings into other languages, “clutch” often becomes an automotive part and “cross-body bag” takes on a more morbid meaning.

And when the company needs to translate product descriptions, machines sometimes spit out impersonal phrases. A warm “Thanks for visiting!” is translated into French as “Merci pour la recherche,” more akin to “Thanks for researching.”

A truer translation

Machine-translation tools are a boon for companies that need to get content like website copy or user manuals into another language quickly and easily. The trouble is that the translations often don’t get the meaning across—so relying on machines alone can lead to text littered with howlers that can confuse customers and potentially lose sales.

Now businesses are searching for ways to get better results. Many use machine translation to get the bulk of their work done but add human oversight to refine the finished product. Others are working on the software itself, trying to teach machines to better understand nuance and context and handle stumbling blocks like social-media shorthand language.

“It’s an eternal problem for machine translation: putting the human element in,” says Ray Flournoy, a group product manager at Etsy whose team works on tools that create smoother translations for its users across the globe.

Machine-translation tools took off in the 1990s as the Internet became more popular, computers got faster and an increasing number of companies pursued global expansion. Even crude translations could help firms open new markets.

Then researchers and companies would write grammatical rules for machine-translation systems. Or, through machine learning and statistics, they could teach the machines to translate using “clean” data from books and other reputable language sources. These tools have advanced significantly in the past decade, but they still can’t match the quality of human translators, experts say.

Still, even those who have complaints about machine translation acknowledge that it does a lot of good—such as making products accessible to more customers, Mr. Flournoy says. He notes that since Etsy began using machine translation live on its site about two years ago, the system has caused fewer than 10 misunderstandings that required Etsy to intervene and refund a purchase.

Etsy uses homegrown language-processing tools alongside Microsoft Corp.’s Translator Hub, which helps companies build custom machine-translation engines. Etsy pulls text from listings that contain problematic words or phrases, has human translators work on them and then feeds the results into Translator Hub. Adding such custom translations helps to train the computers.

Another big area of research: how to translate the messier data that comes from social media and other online sources, where slang, dialect and shortened words like “brb” or “2morrow” can trip up machine translators. Some researchers are building algorithms to help normalize text from social media before it goes through a machine-translation system.

“When content evolved into user-generated content, that’s where machine translation fell short,” says Mudar Yaghi, founder and co-CEO of AppTek, a maker of machine-translation and speech-recognition technology.

Thinking like us

Experts say the next leap forward for machine translation may lie in neural networks, computer systems that roughly try to mimic the way a human brain operates and makes inferences. Based on the data being fed into the translation system, these networks allow machine-learning algorithms to infer a correct translation even if it hasn’t encountered it before.

Many current machine-translation systems can provide hundreds or thousands of possible translations for a given sentence. Statistical algorithms and neural-network technology can quickly rerank those candidates and produce the best translation.

“Translation quality is always climbing the hill to get better,” Mr. Flournoy says.

Source: http://www.wsj.com/