Nlao writeup of Bunescu 2005

This is a review of Bunescu_2005_a_shortest_path_dependency_kernel_for_relation_extraction by user:Nlao.

I enjoy reading this paper very much. It analyzes in detail how syntactic strucures are indication of entity relations.

The method is simple and effective.

Kernel seems to be a good way to handle exponentially many features. However, kernel is against knowledge discovery: making decisions based on examples instead of rules.

[minor points]

- it is good to compare CCG and CFG, but more detailed analysis is appreciated

- having longer paths give larger kernel scores needs more justification. Because we can simply scale the scores so that longer paths generally have same or lower scores.