When computers figure out what you're saying, will they care? Will you?
If the content of your e-mail or a transcript of your phone conversations were being read by someone working for the government, you might consider that a violation of privacy. But most people understand that computer programs and algorithms, such as the Google Adsense program, constantly analyze not only our communications' metadata (whom we spoke to, from where, for how long, etc.), but also, often, the literal content of those exchanges.
What happens to privacy when an algorithm can understand us as well as that government employee might?
At present, there is no computer program that can perfectly interpret human speech, although attempts at solving this problem have spanned five decades. Humans can intuitively comprehend the difference in meaning between words; computers cannot. Several elements factor into determining a sentence's meaning-context, syntax, and logic, for instance, all help us to understand what is being conveyed.
Past attempts to help computers better understand language involved manually coding definitions of words onto disc or memory. This method has yielded little by way of concrete results for the simple reason that language is a lot more complex than mere dictionary definitions.
"I think it's fair to say that this hasn't been successful. There are just too many little things that humans know," says
Programming dictionary definitions is a challenge. Definitions are not always clear-cut, and variation in definitions from one dictionary to another adds to this problem. The solution, the researchers hypothesized, was found not in dictionary definitions but in paraphrasing and the use of more flexible definitions.
Erk and her colleagues combined two distinct approaches to attack the problem. The first piece was Montague grammar (named after philosopher and mathematician
"We use a gigantic 10,000-dimentional space with all these different points for each word to predict paraphrases," explains Erk. "If I give you a sentence such as, 'This is a bright child,' the model can tell you automatically what are good paraphrases ('an intelligent child') and what are bad paraphrases ('a glaring child'). This is quite useful in language technology."
The Texas Advanced Computing Center's Longhorn supercomputer used grammar and syntax analysis, word-meaning models, and first-order logic to predict sentences' meanings. The supercomputer correctly predicted meanings with up to 85% accuracy.
The research was funded by the
Most Popular Stories
- Dell Offers Undisclosed Number of Employee Buyouts
- Saab Gets Back into the Game; U.S. Auto Sales Soar
- American Airlines, US Airways Complete Merger
- Authorities Close to Deal with JPMorgan Chase over Madoff Response
- General Dynamics Plans 200 New Jobs in N.M.
- Unemployed Wait as Lawmakers Debate
- U.S. Stocks Rise on Sysco Acquisition
- Apple Activates Customer-Tracking iBeacon
- Tech Giants Call for Controls on Government Snooping
- 2013 Tech Gift Guide: iPad Mini Still Hot; Chromecast a Great Low-Cost Option