Mnduong writeup of Sha & Pereira

This is a review of Sha_2003_shallow_parsing_with_conditional_random_fields by user:mnduong.

• This paper discusses the use of Conditional Random Fields for NP Chunking. It gives a good discussion of the advantages of CRFs over both generative models like HMMs and sequential classification models like MEMMs.
• Compared to Lafferty et al. (2001), it gives an easier-to-understand explanation of the mathematical aspects of CRFs, mainly because it discusses only the special case of linear chain CRFs as opposed to the generic one.
• A minor point that I found confusing is in Section 2, where the authors gives a "more uniform" notation.
• The authors use different optimization algorithms to train the CRFs, instead of the iterative scaling algorithm that was used by Lafferty et al. (2001).
• Experimental results were shown for the task of NP Chunking, in which the model performed almost as well as the best method that was known.
• In general I think the paper was well-written, with ample arguments to support its claims and a thorough evaluation.