Siddharth writeup of subsequence kernel for relation extraction
A review of Bunescu_2006_subsequence_kernels_for_relation_extraction by user:sgopal1
The paper identifies three patterns ( fore-between , between , between-after ) for extracting relations between proteins. They adapt the subsequence kernel to this task. They also apply this method to extract relations from the ACE corpus. They add some more modifications such as introducing a fourth pattern to capture noun-noun compounds. The results seem to indicate improvement in performance. I would recommend that this paper also be added to the list of optional papers
This paper clearly outlines the derivation of the subsequence kernels, and why is the resulting similarity function indeed a kernel. ( and the DP as well )
A general problem with subsequence kernels is that it sums over subsequences which are not discriminative.
