Selen writeup of Sha and Pereira 2003

This is a review of Sha_2003_shallow_parsing_with_conditional_random_fields by user:Selen.

In this paper:

+They apply CRF's for the task of shallow parsing on NP tags

+ their reasoning for using CRF is to combine the best of the two worlds, generative and discriminative methods.

+ Instead of using GIS for training CRF they train it using a preconditioned conjugate gradient method (improving the gradient, while taking previous direction into consideration)

+They get slightly better results than the other methods

What I don't like about this paper:

- Their method is vulnerable to the choice of prior

- Their methods doesnt yield significant improvement

- P value being high doesnt mean that one method is right the other is wrong