Liuliu writeup of Freitag et al.

This is a review of Frietag_2000_Maximum_Entropy_Markov_Models_for_Information_Extraction_and_Segmentation by user:Liuliu.

Two main things of this paper that I like are:

• The motivation is convicing.
  • MEMM enables researchers to integrate more and richer features into the model instead of only tokens
  • MEMM focuses on state transition and treate observations as given variables.
• I like the way they evaluate models: compare four different models (each one is slightly different from each other) and analyze the reasons for their differences. These reasons illustrate the benifit and importance of their proposed methods (e.g., overlappping features is more effective than token level feature, importance of sequential dependencies)

I also like the way they factored state representations, which solves sparseness and improves generality.