Steven Greenberg

Biography
Steve Greenberg works on web strategy for America Online, where he has also managed instant messaging and community products. Previously he managed the engineering and services groups for a startup in the wine industry, and was senior engineer at Epinions.com. He was a Principal Consultant for Netscape Communications where he founded and ran the Knowledge Management practice. After spending an unspecified number years doing unspecified things for unspecified government agencies, he returned to the world with an MS in Computer Information Systems from Boston University. Your bank probably still uses his code to process deposits.

Position Paper
I spent about half of my 20s working in Heidelberg, Germany, and fell completely in love with the place. One of my favorite uses for Flickr, in fact, is to leave a slideshow running of photos tagged with “Heidelberg”. Flickr shows me photos of smiling tourists, mountain vistas, a moss covered castle… and industrial printing presses. It turns out that “Heidelberg” is also the brand name of a company that makes the sort of presses used to produce newspapers and glossy magazines. Figuring out how to separate these contexts is one of the central hurdles facing systems that depend upon user classification.

Flickr has created value by gathering these pictures and helping me find them, but they are limited by the fact that uploaders have little incentive to provide more than the bare minimum context they require for their own use. An enterprise can hire librarians to build a structured taxonomy and force everyone to fully qualify their tags, but it is too much to expect that mass market consumers will learn and follow any set of rules.

This is not to say that I think user tagging is useless or unworkable in the mass market. Rather, I suggest that a change of approach can get us to a “good enough” state without asking an unrealistic amount of effort from either taggers or searchers. I propose that we turn the standard clustering approach on its head. Rather than grouping things that are similar, we should focus on separating what is dissimilar.

When the user assigns a tag, the system could ask one or two follow up questions based upon other tags related to the one just provided, but which are infrequently used together. It could also suggest common synonyms already in use. These actions can help the community converge upon a standard vocabulary while still allowing individuals to expand it when they see fit. Thus, the Flickr user who posts a photo simply tagged “Heidelberg” would be asked whether “castle” or “printing press” is more likely to describe the content. A learning system could identify sets of tags frequently and infrequently used together, relying upon a base taxonomy to provide suggestions in cases where there are not enough existing tag sets to make statistical predictions. By using word frequency, this approach can also help classify text documents.

The ambiguity of human language is what makes it such a flexible tool. That same ambiguity, however, demands that real communication be based upon either voluminous detail from the speaker or a feedback loop. Simulating a feedback loop with follow up questions gets us good enough results with the minimum of effort.

Back to Social Computing Symposium 2005