16 May 2005

The alternative to emergent ontologies

I recently summarized a few thoughts about merging the top-down and bottom-up systems of WordNet and tagging as a good solution to fixing the brittleness of a tagged ontology. Clay Shirky suggests that combinatorial principles [Wikipedia] could be used to find the ontology inherent in the set of tags. Instead of merging a semantic hierarchy with search engine predicate calculus, the sematics would be derived from the liklihood that tags occur in specific combinations.

In natural language processing [Wikipedia], many parts of speech taggers will use a similar method (as a hidden Markov model) to tag unknown words based on the surrounding, known words.

It's a compelling argument, and I should have considered it.

Still, if the comparison can be continued, there are no completely probablistic POS taggers. They at least will know about syntactic items such as articles or common morphemes in order to tag the unknown words. It's a popular method, though, and the principle could be useful for tagged ontologies.

[ posted by sstrader on 16 May 2005 at 3:43:30 PM in Programming ]