KeisukeKamataki writeup of Cohn 2005

This is a review of Cohn_2005_semantic_role_labelling_with_tree_conditional_random_fields by user:KeisukeKamataki.

Summary: They applied CRF for the nodes for syntactic parse tree to do semantic role labeling. Given a parse tree, many different types of features were extracted (basic features, context features, common ancestor of te verb, feature conjunctions, default features, and joint features). The value of F-measure was about 70.0% and looks a kind of accurate although there was a high variance of performance for different types of semantic role relations when we see the detailed performance.

I didn't like/didn't well understand: Although this paper is very clear about what they tried and achieved, more detailed analysis for the experimental result might be needed. Especially, I'm interested in why there is such a large variance in performance in different semantic role. More detailed error analysis to explore this issue might make this paper all the better.