Put comments here for "A Note on the Unification of Information Extraction and Data Mining using Conditional-Probability, Relational Models" by McCallum and Jensen:
From James Cussens: This seems the way to go if you can get it working! I like the fact that uncertainty from one step (IE) persists into the next step (DM) - or rather that there is a merging of these steps. But won't these CRFs be massive? In your closing section, you state that relational data are straightforwardly modelled in undirected graphical models by using tied parameters. Presumably it's the distribution "behind" the data rather than the data themselves that are so modelled, but that's a pedantic point. More interestingly, you're claiming to have a relational model which does not involve an explicit representation of the (related) objects. I think there's a lot to be said for such an approach, primarily on grounds of parsimony - the simpler our models the better.