Teaching computers the meaning of words
Teaching computers to understand how we feel as transmitted by our voice and the contextual meaning of words is a challenge researchers are trying to solve.
"People say the damnedest things in the damnedest ways," says Google research director Fernando Pereira in describing the challenge of computers understanding human speech. Computer language, after all, is absolute, where the human equivalent is notoriously messy and haphazard.
We understand meaning on the fly because we're wired for contextual patterns, to deduce meaning, but a search algorithm might respond to a query, for example, about a robber being charged in court by looking for crimes in the wrong places. Now, efforts to teach machines about context rather than just word meaning might change that.
When it comes to teaching a computer context – the associations between words – University of Texas linguist and statistician Katrin Erk has designed a theoretical framework that gives a computer clues about what a word is likely to mean based on common relationships with other words. She and her team then gave the system 100 million words from literature to crunch and let it loose to discern likely meanings.
While talking to a computer isn't the goal of Erk's research - "We test on how well we match the conclusions people draw," she says - more words and faster computer power raise the startling possibility of literally speaking with a machine in our own conversational language in real time.
Methods such as Erk's are just one part of Google's strategy, with Pereira saying the company is using several "teaching methods" to program machine understanding – including contextual relationships.
"What is to granddaughter like brother is to sister?" Pereira says. "We can teach the computer that by mining a lot of data and graphing a 'neural network'."
The applications go beyond just typing or dictating a question into a search engine – we could end up with a Siri-esque software agent that seems more alive than ever, a capability Speaktoit's Gelfenbeyn thinks will soon be critical.
"As interfaces expand to our cars, offices and homes, natural language understanding by these inanimate objects is a must," he says.
Even the emotional inflection of human communication is going digital. Systems such as Vivotext let you program the emotions of pitch and timbre in speech. Vivotext vice president of strategy and business development Ben Feibleman says doing so in the other direction – having the computer account for them – wouldn't be difficult with automated pitch detection technology.
Together with different languages, accents and region-specific linguistic idioms, it all adds up to a huge amount of processing to be done. But Ilya Gelfenbeyn, chief executive of Speaktoit, the company behind a popular Android alternative to Apple's Siri, can see light at the end of the tunnel.
"It's no longer a pipe dream," Gelfenbeyn says. "We're getting much closer to widespread use across a range of devices."
According to Google, a Star Trek computer is the long-term vision of the company. So when will we get there? "It depends on your expectations," Pereira says. "Five years ago I couldn't even conceive of the things we can do today. What we can do in voice search now is better than a year ago and will be better again in year."
Of course, one of the biggest questions will be who pioneers or buys such capability. Erk hasn't had any offers from the private sector yet, but the project is still at prototype stage and she says there's a lot of work to be done bringing several applications, some of them third party, together seamlessly.
Google's Pereira confirms he's aware of Erk's work, and when asked if Google's efforts will come from engineering internally or acquisitions, he points to work already being done in-house along similar lines, but adds: "Not all smart people work at Google, to the extent we can license their technology if it makes business and technological sense, we're very open to it."