[Freebase-discuss] Question on related topics
Paul Houle
paul at ontology2.com
Thu Aug 4 21:08:26 UTC 2011
I'll just add that the problem of finding a good list of "topics
related to X" in Freebase takes some effort.
I think the main strategy that Freebase uses to make a list of
things that share a type (filtered somehow) with the topic that are
ranked by some kind of importance score. Something like this is used to
generate some of the navigation on the top page of Ookaboo, but it
doesn't scratch the related topic itch.
Ookaboo tries to find related topics for things by processing link
relationships from Wikipedia. You could try the same thing following
some other semantic network. Recall is easy, precision is hard, and
I'm not happy with the results. It's not hard to make an importance
score for topics (think PageRank) but importance scoring for
relationships is hard. You've got to consider different types of
relationships ("Is it more important that A was in movie B or more
important that they were born in city C?" And also, if some guy was in
20 movies, which one matters the most? The huge blockbuster that he
played a bit part in? Or the low budget movie that he starred in?
seevl.net actually has something that works well, but it's for a
particular domain. It's processing the semantic relationship network,
using valuable data chosen from both Wikipedia and Freebase: almost
certainly there's some kind of ontology specific knowledge and a scoring
function involved. Extending this to "everything" might involve a
substantial effort to develop a rulebase, but it would be sweet to find
some pixie dust that makes it unnecessary.
More information about the Freebase-discuss
mailing list