Selen writeup of Cohn 2005

This is a review of Cohn_2005_semantic_role_labelling_with_tree_conditional_random_fields by user:Selen.

In this paper they define an undirected random field over the sentence's parse tree for semantic role labeling. They prune the parse tree so that only nodes that are siblings to the node remains, and they add children of the prepositional phrases to the model, if there is any. Their feature set is not so different from the earlier work.

What I like about this paper, is that their approach is different and could be promising. What I don't like is that they do not compare their method with the baseline, and they don't give a reasoning why they include children nodes of prepositional phrases. Also it would be better if they didn't prune the tree (and automatically prune it using shrinkage methods etc)