Nschneid writeup of Sha 2003

This is Nschneid's review of Sha_2003_shallow_parsing_with_conditional_random_fields

CRFs for NP chunking with BIO tags. Compares several training methods.

I like this paper's overview of generative sequence models vs. local classification models, and argument that CRFs give the best of both worlds. Their formal description of CRFs is very clear. I also like that it compares several training methods, including runtime and convergence rate.