Bbd writeup of shortest path dependency kernel

This is a review of Bunescu_2005_a_shortest_path_dependency_kernel_for_relation_extraction by user:Bbd.

This paper gives a kernel trick based on the assumption that relation between 2 entities can be defined using the shortest path that connects them in the dependency graph. They study using this technique with CFG which models local dependencies very well and CCG which models local as well as long distance dependencies. Because dependency graph is always connected they are bound to find shortest path each time. The simple kernel that they used computed number of common features between 2 entities.

I liked the way they learn the set of features using SVM in efficient way. The number of features is cartesian product of word classes so manageable sized. Also kernel computation time is in order of (length of path), which can be restricted.

I didnt quite understand why CFG based version always outperforms CCG based version.