A new search engine from DolphinSearch Inc. is designed to emulate the intelligence of its namesake.
Dolphins use pattern recognition to identify objects, according to Professor Herbert L. Roitblat, chief science officer of DolphinSearch, who researched dolphin cognition for more than 15 years.
Roitblat created a computer model of the powerful pattern recognition portion of a dolphin’s brain and has applied the model to the problem of understanding written human language.
“I’m a weird person,” Roitblat said. “I was interested in how people understand ambiguous words.”
DolphinSearch is based on Roitblat’s research. It enables computers to extract suggested meaning from written language. The Ojai, Calif., company claims its search engine can retrieve more relevant results because it recognizes the actual content of documents.
For example, Roitblat studied the word “strike.” He said people do not notice the ambiguity of it.
“You can strike a pose, you can strike out, you can strike out for Chicago, an idea can strike you, and you can strike a match,” Roitblat said. “The reason you don’t notice how many there are is because whenever you use it in a sentence, within 200 milliseconds you understand what I have in mind.”
It then occurred to Roitblat that neural networks could also be used to model how people understand those ambiguous words. He also realized that this type of ambiguity is a problem on the Internet.
“When you enter a word in the search engine, the search engine doesn’t know which meaning you have in mind, so it gives you all of them,” Roitblat said. “Neural networks could be used to train for community on a particular meaning.”
Andy Kraftsow, president of DolphinSearch, said the product is a pattern-recognition mechanism, like a mammalian brain.
“(It) understands what something is about in pretty much a matter that is indistinguishable from the way a human recognizes (information),” Kraftsow said.
Roitblat said that it is all a matter of pattern recognition. “When a dolphin echo locates, it sends out a sound through its forehead, a very high frequency, broadband sound, that hits an object and reflects back. What the dolphin responds to is a profile of all the different frequencies of the sounds that are contained in the echo.
“If we have a set of documents of known relevance, then we can train on those,” Roitblat said. “Our neural network shows how [a] community uses the words, and it knows what meanings people attach to them.”
But DolphinSearch’s capabilities may not be unique, according to one researcher.
Henry Lieberman, a research scientist at the Media Lab at the Massachusetts Institute of Technology, said most search engines recognize the problem of narrowing down results. They perform statistical analyses of popularity, count the frequency of words or rely on analysis by human experts for relevance.
Lieberman said his technical opinion of DolphinSearch is that it’s a sound, reasonable offering and that it “seemed to do better, on average” than other offerings.
“I wouldn’t say it’s earth-shattering. I think the theory of it is, in general, good.”
DolphinSearch may face direct competition from a technology called latent semantic indexing, Lieberman added. Autonomy, a British company, categorizes material using neural networks, pattern matching and a technique dubbed Bayesian analysis.
A subset of DolphinSearch, called KnowledgeBox, could help corporations retrieve information from an Intranet, Roitblat said “Almost everybody acknowledges that they can’t find stuff. KnowledgeBox will do that.”
KnowledgeBox is an automated plug-and-play network retrieval appliance. It reads documents in a corporation’s network, understands what they are about, recognizes their point of view, and remembers where they are located.
A DolphinSearch (www.dolphinsearch.com) implementation costs approximately US$10,000. DolphinSearch can be reached at (805) 640-9984.
With files from IDG News Service.