Liuy writesup Bunescu 2006 subsequence kernels

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

The paper exploits generalize subsequence kernels for relation extraction task. Three types of subsequence patterns are used to describe the relationships between two entities. Compared with the rule-based systems, their approach shows some advantage on a protein interaction dataset. Compared with the dependency tree, their approach also better extract top-level relations on the ACE corpus.

I like the paper because they not only show us some empirical advantage on one or two datasets, but also tell us why. The fact that unbalanced distribution might harm the recall score, can explain the benefits of using the relation kernel cascadedly.

However, I do not like the parameter tuning in an adhoc fashion. The principles on tuning the parameters should be given. I think the mutual relations between NER with relation extraction are worthy further studying. Also, the computation of relation kernel involves the counting of common subsequences. Though they already try to save the amount of counting work, by separating these common counts calculation, the computation complexity still need to be improved.