Apappu writeup of Freitag et al.

This is a review of the paper Frietag_2000_Maximum_Entropy_Markov_Models_for_Information_Extraction_and_Segmentation by user:apappu.

• This paper talks about combining the advantages of HMMs and Maximum Entropy models.

• Authors state two problems with traditional approach (of HMM):

 * not addressing the interdependency between features
 * emphasis on maximizing the likelihood of the observation sequence when task is conditional in nature (state sequence given observation sequence). 

• In one of the variant approaches suggested by authors: observations in states instead of transitions approach has an interesting advantage of reducing the number of parameters and project it useful for applications when training data is sparse.

• I also liked how it was shown that structure regularities actually matters besides line features. (refers to Table 4)