Liuy writeup of Cohn 2005

From Cohen Courses
Jump to navigationJump to search

This is a review of Cohn_2005_semantic_role_labelling_with_tree_conditional_random_fields by user:Liuy.


The paper explores a tree structure and use efficient CRF inference algorithms for the semantic role labelling. Their approach is able to model parent- child interactions and outperforms maximum entropy classifier.

I like the work for the following reasons. It shows that structure like sequence of words or chunks is not as effective as the parse constituent structure. Further it defines a random field over a syntactic parse tree for each sentence and predict a semantic role label for nodes in the tree. Conducting the task of constituent selection and labeling at the same time, their approach has better generalization performance on the WSJ test than maximum entropy classifiers.