Mnduong writeup of Cohn & Blunsom

This is a review of Cohn_2005_semantic_role_labelling_with_tree_conditional_random_fields by user:mnduong.

• This paper introduces a method for semantic role labeling using conditional random fields. Examples are represented as pruned syntactic parse trees, together with each node's role label. Nodes are kept in the tree according to several rules, e.g. words at the leaves and their syntactic categories are not included, and only siblings to a node from the verb to the root are kept.
• Features are defined over cliques in this tree - either one node cliques (tree vertices) or two node cliques (tree edges). The CRF was trained using limited memory variable metric.
• The paper gave experimental results but did not provide any comparative evaluation against other methods. "Preliminary" experiments were carried out with only one-node clique features, which yielded poor results. This was used to imply that the CRF would outperform standard maximum entropy classifiers. However, no concrete results were provided, perhaps due to space limitation.