Yandongl writeup of Cohn 2005

This is a review of Cohn_2005_semantic_role_labelling_with_tree_conditional_random_fields by user:Yandongl.

This paper explored the possibility of using CRF for SRL. Usually CRF is defined on some chain-linear model while it's hard to convert the parse tree structure into some linear model.

Features include single and pari-wise ones which describe the interaction between parents and children. Training approach is LMVM, which is new to me.

Maximum entropy model has been used a lot for SRL tasks. But seems no comparison is reported between CRF and MEM?