Bbd writeup of subsequence kernals

A review of Bunescu_2006_subsequence_kernels_for_relation_extraction by user:Bbd

This paper presents a subsequence kernel technique that works with sequences containing combinations of words and word classes. They base their approach using 3 kinds of features : Fore-Between, Between and Between-After. To compute distance between 2 entities s & t relational kernel computes number of common patters between 2 sentences which are restricted by the 3 types.

I liked the way they tried to generalize the system. They report good results on biological as well as newsgroup data. Also I liked the way they train one binary SVM for relation detection and train second multi-class SVM to do relation classification.