Thursday, June 23, 2005

ScentHighlights and "spreading activation" ... Google searches, apparently, are so twentieth century. The CSMonitor has a story today on what's next:

The reading experience online "should be better than on paper," Chi says. He's part of a group at PARC developing what it calls ScentHighlights, which uses artificial intelligence to go beyond highlighting your search words in a text. It also highlights whole sections of text it determines you should pay special attention to, as well as other words or phrases that it predicts you'll be interested in. "Techniques like ScentHighlights are offering the kind of reading that's above and beyond what paper can offer," Chi says....

ScentHighlights gets its name from a theory that proposes that people forage for information much in the same way that animals forage in the wild. "Certain plants emit a scent in order to attract birds and bees to come to them," Chi says. ScentHighlights uncovers the "scent" that bits of information give off and attract readers to it.

If the reader types in "Wimbledon tennis," for example, ScentHighlights would highlight each word in its own color in the text, as search programs do. But ScentHighlights adds additional keywords in gray that the system has inferred that the reader would be interested in (perhaps "US Open" or "Andy Roddick"). It would also highlight in yellow entire sentences that it deems likely to be especially relevant.

To do this, ScentHighlights combines two approaches, noticing how often words are near each other in text and using a technique called "spreading activation." Chi says: "It basically mimics how humans retrieve information." ScentHighlights actually knows nothing about tennis, he says. "It's a purely statistically based technique."

Off the top of my head, the most interesting thing about this is that it could help document media bias. Imagine doing a search on "evangelical", for example. The keywords ScentHighlights would come up with would probably include "socially conservative" or "religious right". By contrast, if you searched "liberal agenda" you'd quite possibly come up with something like "unpatriotic". By matching words with qualifiers and attendant phrases that have a far higher frequency than would be statistically expected, the program could help indicate just how pervasive media bias truly is.

Unfortunately, ScentHighlights hasn't quite hit the market yet. But there is at least Amazon's Statistically Improbable Phrases, which turns up some pretty funky stuff within books themselves.


Post a Comment

<< Home