KeisukeKamataki writeup of Bunescu 2005

This is a review of Bunescu_2005_a_shortest_path_dependency_kernel_for_relation_extraction by user:KeisukeKamataki.

Summary: In order to do relation extraction between the two named entities in the dependency graph, they used a kernel which is based on the shortest path between them. The shortest path is computed from a dependency tree of a sentence taking into account of its dependency structure. They trained SVM and applied it to ACE corpus for 2 scenarios. One is the simply to predict 5 categories (ROLE, PART, LOCATED, NEAR and SOCIAL) and the other is to divide the classification into 2 stages; relation detection/relation classification. For first, scenario, CFG-parse(suited for local dependencies extraction) based method significantly out performed K4 (a sum between a bag-of-words kernel and the dependency kernel) consistently especially in recall. In second scenario, CFG-parser based one kind of improved K4 except for in precision.

I like: This paper well describes how to incorporate syntactic information with kernels and classification. It might be even better if they explored why there is relatively smaller performance improvement from scenario 1 to scenario 2 with their kernel.