Bbd writeup of semantic role labelling with tree CRF

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:Bbd.

This paper applied CRF on parse trees of sentences to do semantic role labeling. They define SRL as identifying which group of words act as arguments to a given predicate.

They model SRL as random field over the parse tree and perform both constituent selection and labeling simultaneously. Since parse tree is acyclic graph, exact inference is possible using CRF and can be done just one pass. They use features like context features, common ancestor, joint features which were used in past by other people.

In all the paper selects some existing tools developed by other paper and applies it to SRL problem to give efficient training time. They claim it to be better than maximum entropy technique used earlier for SRL by allowing parent-children interactions to be accurately modeled. Also leading to the fact that modeling random field over parse tree is better than modeling data as sequence of words/chunks.

The experimental results are not quite expressive. I couldn't figure out how they claim to outperform comparative methods.