Rbosaghz writesup Bunescu 2006 subsequence kernels

This will be a review of Bunescu_2006_subsequence_kernels_for_relation_extraction by user:Rbosaghz.

This paper is an example of how picking a good kernel can aid learning in the same way that picking good features can. In this case, the learning task is relation extraction. What I particularly liked about this paper is that their computation of the kernel is a multistep process which uses theorems from learning, governing how kernels can be combined to create new kernels (sums, products, etc). I had not seen these theorems used so heavily in practice, and it's nice to see them really ensuring the validity of the kernel while providing better experimental results.