Nschneid writeup of Cohn 2005

This is Nschneid's review of Cohn_2005_semantic_role_labelling_with_tree_conditional_random_fields

This paper uses a (non–linear-chain) CRF over constituency parse tree nodes/edges in order to do PropBank SRL. The predicates are observed. They show this does better than a maxent model (i.e. without any pairwise parent-child features). Evaluation is on WSJ and Brown; unfortunately they do not compare to other work. Test set F1 is 71.2 (how does this compare with a state-of-the-art system?).