Bbd writeup of Fei Sha 2003

This is a review of Sha_2003_shallow_parsing_with_conditional_random_fields by user:bbd.

This paper presents a technique to train CRF model for shallow parsing CoNLL task. They claim to have performance comparable to any base noun phrase chunking method and better than any single model. The paper very well describes how CRF brings advantage of discriminative models(leveraging statistically related features of input) and generative models(finding globally optimal solution sequence). I liked the way they avoid overfitting by introducing spherical Gaussian weight prior for penalizing Likelihood. They have done extensive experimental comparison between various shallow parsing methods, but the comparison metrics they use like McNemar's test seem fuzzy and can;t convince how their model will perform better than others in other problems like NER and POS tagging.