Siddharth writeup of cohn tree conditional random fields

This is a review of Cohn_2005_semantic_role_labelling_with_tree_conditional_random_fields by user:sgopal1.

This paper applies CRFs for Semantic Role Labeling. The input is a bunch parse trees and the semantic roles of the words/phrases are labeled. They find that using a duplication leads to over-prediction when using CRF's. They incorporate both vertex level and edge level features into the CRF's. There seem to be rich collection features representing information about the structure as well linguistics.

There seems to be no comparative evaluation. Would have been better if they have atleast reported the results based on pure vertex level features.