Siddharth writeup of shortest path dependency kernels

This is a review of the paper Bunescu_2005_a_shortest_path_dependency_kernel_for_relation_extraction by user:sgopal1.

This assumption of this paper is that the shortest path between two entities in the dependency graph captures the relation between them. They give a few examples of dependencies and how the shortest path in the dependency graph captures it. Each word is associated with POS tags,word-classes etc. They define a dot-product space by mapping each path in the graph to a higher-dimensional vector. The vector is a sequence of nodes in the shortest path, and each node is represented a smaller K-dimensional vector of its components such as POS-tags, word classes etc. In their experiments they have a total of 24 relation types. They compare their methods ( 2 variants with different grammer CCG, CFG ) with a convolution tree kernel. They also perform the experiments using a multi-class SVM as well as OVA SVM. The second set of results (S2) doesnt sound too convincing, it would be nice to see to a balanced increase in both precision and recall than an unbalanced one.