Rbalasub writeup of Cohn and Blunsom

A review of Cohn_2005_semantic_role_labelling_with_tree_conditional_random_fields by user:rbalasub

The authors use CRFs for semantic role labeling. Since it's awkward to express SRL as a linear chain labeling task, the authors frame the problem as a tree node labeling problem in a parse tree. They explore two heuristics to prune the tree before attempting to label it. Since CRFs based on trees need more involved techniques for decoding than linear chain CRFs, belief propagation is used.

It would have been interested to see a comparison of tree CRFS vs chain CRFs to see the advantage that a deeper structure provides. Parsing itself is not a 100% foolproof process and it will be interesting to have an idea of when/if it's worth paying the price of parsing