Bbd writeup of Frietag 2000

This is a review of Frietag_2000_Maximum_Entropy_Markov_Models_for_Information_Extraction_and_Segmentation by user:bbd.

The paper presents a Maximum Entropy Markiv Model (MEMM) to predict best state sequence from observation sequence. The basic difference from HMM lies in combining the transition and observation functions in HMM models by a probability function of current state given last state and current observation.

I liked the detailed evaluation done on 4 models : ME-stateless, TokenHMM, FweatureHMM & MEMM. I also appreciate their argument of high precision scores of MEMM being very important for practical applications like question answering system, where the system can be almost automated using high precision MEMM.