Intelligent Agents Come of Age
by Chris Feola
Intelligent agents have become the videophones of the '90s.
Admit it--you (and everyone else) fell in love with videophones the first time you saw them, some time after Dick Tracy first strapped one on.
Yes, videophones were one of the coolest ideas out there. Too bad they were a bigger flop than "Howard the Duck."
Like videophones, intelligent agents have captivated imaginations. It's a cyberdream--an intelligent friend who knows exactly what you want (and what you don't) and roams the 'Net, sifting and collecting information for you. You know, sort of a virtual gopher.
In practice, intelligent agents have proven virtually gothic.
It's not hard to understand why. Most current intelligent agents involve some combination of Byzantine forms, do-it-yourself start pages and search-engine technology.
The bottom line is that computers are as stupid as a box of rocks. Oh, they're good enough at finding whether a particular word is or is not in a document. It's just that computers generally find it almost impossible to place that word in context.
Think of it this way: Let's say you're going to put together the rules for the computer to use when dealing with the word "love."
OK, done yet? That was easy, right?
Of course, you remembered to include rules for "love" to differentiate between the way you feel about your parents, your spouse, your child, your pets, your car, your friends, your job, your fave outfit.
But did you also include rules for "love" when it is used sarcastically? Ironically? As a metaphor?
You get the idea. There are simply too many possibilities for any computer to calculate.
So Michael Lynch narrows the possibilities.
The founder of Autonomy Inc. of Palo Alto, Calif., helped develop the basis for a different type of agent technology while earning his Ph.D. at Cambridge.
"The bit that matters is the ability to take a document and work out the ideas being talked about," says Lynch. "The same idea can be talked about very differently. Take the London newspapers, the Sun and the Times. Both of those papers cover the news, and often the same stories, but with different words and in different ways.
"The approach we take is a pattern-based approach. It's like matching fingerprints, or spotting fraud," says Lynch.
Autonomy's Agentware line, released in August, uses this approach to develop probabilities about context. The line includes a content server, a personal-content delivery system, a push component, user profiling, a community-building application, a targeted blocking module and a user filter.
For example, says Lynch, take the word "Apollo." If the word "Apollo" appears in the same paragraph with the word "rocket," then there is a high probability that the paragraph is about the American space program. If the word "Apollo" appears with the word "temple," then it is quite possible that the topic is Ancient Greece. If it appears with the word "theater," then there is a high probability that the topic is Harlem or African American culture.
Notice that in each of these instances, the context--the topic--is not a word that appears in the text.
Clearly, better agent technology is important to newspapers and their online products. But here's even better news: Agentware can build all the links for your Web site on the fly--and both individual links for users based on their reading patterns and links to other branded content from your organization.
"As someone reads the documents, Agentware can work out what the person is interested in," says Lynch. "Let's say the first page the guy sees has an article on beer." Agentware watches the user navigate, then uses its context analysis and probability prowess to provide links for the reader. "Now he's navigating the content, not by going though complex navigation, but by six or so links. And we can provide side links to other branded products--from the Sun to the Times version. We have a way of passing people around the syndicate.
"The reader just wants info on beer," says Lynch. "He doesn't want to deal with all this other stuff," like the Boolean math required for effective searching on most of the 'Net.
So, what about the two obvious questions: How well does all this work, and can it handle an operation the size of yours? Very well, and it handles LineOne, the vast site that is a joint venture between Rupert Murdoch's News Corp. and British Telecom. The site includes content from Murdoch's array of British newspapers and broadcast outlets. You can check out LineOne and see an Agentware demo by pointing your browser to Autonomy's Web site (http://www.agentware.com).
Chris Feola is director of the media center at the American Press Institute, Reston, Va. E-mail, feola@apireston.org; phone, (703) 715-3333.