|
I'm interested in applying a CRF model of the form P(y|x) = exp(-E(x,y)) when the dependence between Boolean-valued y variables includes pairwise interactions, i.e. E(x,y) = sum_i Q_i(x)y_i + sum_{i,j} Q_{i,j}(x) y_i y_j. For any x I can approximately minimize E(x,y) for the optimal y (or evaluate required expected values) up to about 500 y values. Of course this inference is very hard in general, but with specialized hardware I can do a very good job very quickly. I have tried this on a few structured classification problems (scene, reuters), and I'm looking for more examples. Can anyone suggest problems and training sets where I could apply such a CRF? Ideally, I'd like training sets where the features had already been generated, and where we expect pairwise interaction in output labels y_i y_j to be important in successful generalization. |
