Artificial intelligence tries to sort out the Babel of languages
Every working day an army of 1,500 Canadian translators converts English documents to French, and French to English - about a million words a day. But if Canadian officials have their way, computers with artificial intelligence will pick up the slack so that Canada isn't buried by a tidal wave of words.
Computer translation has been around for decades, but it was not sophisticated enough to be of much use until the mid-1970s. Now, however, artificial intelligence techniques are making computer programs better able to sort the complexities and subtle distinctions of language needed for efficient translation.
Already Canada is spending $90 million annually to translate ``the equivalent of a Bible and a half'' every day, says Alain Landry, assistant under secretary of state for official languages and translation. And he expects the demand to grow.
Canada is, of course, in a unique position since a 1969 law requires translation of official documents into both English and French. Yet other countries, particularly those in the European Community, sorely need a more efficient way to deal with translation of documents.
The European Economic Commission estimates that translation of documents currently costs $10 billion to $15 billion a year worldwide. Europe as a whole is home to 700 million people speaking more than 30 languages.
As much as 60 percent of the 12-nation European Community's budget is spent on translations of its nine official languages. Business would flow considerably faster but for the expense and necessity of translation.
``We are now a global village and we need to communicate like one,'' says Astrid Johanson of the American Translators Association's computer translation committee. ``Human translators are not going to become extinct, but we need better machine translation to help them.''
Despite its promise, computer translation still has only recently ``graduated from infancy to being a juvenile,'' Ms. Johanson says.
In order to be of real help to a translator, a computer program must provide translations that are 70 percent accurate or better. But one problem all computer programs stumble over is the homograph - words that are spelled the same, but mean different things.
The word ``float,'' for example, can be used as a noun or a verb. A banker and a naval architect could each use the word in a different way, rendering a typical computer program unable to discern which French or German word for ``float'' to use.
But Jonathan Cave, director of marketing for Logos, a Dedham, Mass., computer software company specializing in translation programs, thinks his company may have the problem in hand. It is Logos's program that was chosen by the Canadians for a pilot program.
``We've given our program the ability to evaluate the syntax of the sentence, the parts of speech,'' Mr. Cave says. ``Then it layers on real-world knowledge [through artificial intelligence techniques]. It knows, for example, that a rock can't fly - unless it's flying through the air.''
While machine translation involves computers that have built-in dictionaries, spell-checkers, and other tools to assist a human translator working at a computer terminal, some are more sophisticated and work automatically.
The problem is that ``automatic translation'' still requires a human to proofread for nuance, and the accuracy rate is still below 70 percent with many programs - sometimes making them more of a hindrance than a help - depending on the complexity of the document.
The new software developed by Logos that is being tested in the Canadian pilot program is ``fully automatic.'' That means the documents are processed at night by a computer, and a human translator goes over them by day. Cave says this increases translator efficiency from 300 to 800 percent.
The market for artificial intelligence ``natural language'' programs is small (about $20 million) but could grow to $500 million by the end of the decade, says Dataquest, a San Jose, Calif., research company.
