Ksuravar writeup of Frietag et. al.
From Cohen Courses
Jump to navigationJump to searchThis is review of Frietag_2000_Maximum_Entropy_Markov_Models_for_Information_Extraction_and_Segmentation by user:ksuravar.
Summary:
- The paper presents new markovian model MEMM
- The paper uses conditional probabilities of states given observation sequences (and previous state) as apposed to probability of a state given the previous state and probability of the observation given the state independently.
- The observations themselves can be dependent
Comments:
- Though the paper mentions in the introduction that "GIS (used for the parameter estimation in this case) is similar in form and computational cost to expectation-maximization alogorithm" I think providing the running time or some measure where we can compare the complexity the new model whoud be great. THe reason is I believe new model is computationally more expense then old and before appreciating the results we should also think about the computational time it takes for giving those results.
- One think I liked about the paper is that they not only gave the new markovian model but also considered the "three classical problems" of markovian model and gave the variant of 'forward-backward' procedure.