The days of madly flipping through bilingual phrase books while trying to converse with somebody in another language is fast giving way to a technological alternative developed in the Bay Area and elsewhere.
Armed with software from Google (GOOG) and other companies, computer users can communicate in dozens of the 7,000 languages the United Nations says are spoken worldwide and translate entire Web pages with the push of a button. They also can use their smartphones while abroad to chat with people in their native tongues and even use the phone's camera to decipher local restaurants' menus.
But while these so-called machine translation systems have made remarkable progress in recent years, experts caution that they remain severely error prone.
"If you are a tourist on a street needing to know where the train station is, these tools are pretty good," said Elizabeth Bernhardt, director of Stanford's Language Center. However, "if I were in a business context and I had some serious negotiations going on, I would not rely on machine translation." Besides, establishing rapport through a gadget is tough, she said, because it's "like talking through a Teletype."
Using machines to sort though the global babel of languages has long been a dream of scientists. Georgetown University and IBM conducted the first public demonstration of the idea in 1954, converting 49 Russian sentences into English. Yet working out kinks in the concept hasn't been easy.
For years, the usual method for performing automated translations was to cram the computerized systems with bilingual dictionaries and detailed grammatical rules prescribing how sentences in different languages should be structured. But accounting for the vast number of rules and exceptions is hard. So Google uses what is termed statistical machine translation.
Asked to change a French sentence into English, for example, that system scours a huge Google database of previously translated French expressions, compiled from various Internet sources. Then, based on its statistical analysis of the data, it responds with the most probable answer -- all in the blink of an eye.
Claiming 200 million active users a month, Google Translate is continually adding new languages and currently handles 66, ranging from Arabic, Norwegian and Hungarian to Urdu, Telugu and Swahili. Users simply type in a phrase and tell it what language they want it changed to.
Google's Chrome Web browser makes it just as easy to translate foreign-language Internet sites. When someone happens onto one, they can click on a bar that turns the pages into the language they prefer. The feature is so useful that the California Academy of Sciences and city of San Francisco have built it into their home pages.
Google Translate also is available as a mobile app. That way a U.S. traveler trying to converse with someone in Vietnam, for example, can speak English into their smartphone and it will translate their words aloud into Vietnamese. The Vietnamese person's reply is translated the same way back into English.
A separate feature lets a person translate text with their phone's camera. If a U.S. traveler is stumped by an entree listed on a restaurant's menu in China, they can snap a picture of the item, brush their finger over it and their phone will tell them what it is in English.
Similar features are offered by Microsoft's Bing Translator and Israeli software provider Babylon. But experts caution that such systems aren't foolproof. When this newspaper tested how well they translated several Spanish phrases into English, the results varied widely and some were badly garbled.
Google software engineer Josh Estelle noted that some languages, such as Spanish, are full of male and female linguistic nuances -- as well as formal and informal ways of addressing people -- that "can be a big challenge for us." Different dialects can be tricky, too.
"There is a very clear distinction between Brazilian Portuguese and the Portuguese they speak in Portugal," he said. "Our system doesn't distinguish between that. We wish it could."
Menlo Park research outfit, SRI International, is among those working to improve the technology. One idea it is studying is to give machines the ability to spot linguistic ambiguities that could cause translation errors, prompting it to seek more information from the person making the query.
In other words, the machine would "know when it's making a mistake and ask for clarification," according to Kristin Precoda, director of SRI's Speech, Technology and Research Lab.
Yet Sonia Wichmann, president of the Northern California Translator Association, believes humans will always provide the most reliable translations and warns against ever taking a machine too literally.
"You have to really watch out," she said. Otherwise, "you could easily get yourself into a lot of trouble."
Contact Steve Johnson at 408-920-5043. Follow him at Twitter.com/steveatmercnews.
Machines don't always agree when translating foreign languages, especially with idiomatic expressions. Consider the Spanish saying "A caballo regalado no le mires el diente," which means, "Don't look a gift horse in the mouth," according to Cassell's Contemporary Spanish handbook. Here's how three popular Web-based services turned that into English:
"Beggars cannot be choosers"
"Tooth don't look a gift horse"
Babylon: "Gift horse you not look at the tooth"