Liuy writeup of Frietag 2000

This is a review of Frietag 2000 Maximum Entropy Markov Models for Information Extraction and Segmentation by user:Liuy

This paper tries to combine HMM and maxent in a more general model, that connects state transition probabilities to features of the sequences that we are interested in. The maxent is used to train exponential models as a representation of state transition probability in a HMM model. I like the paper in its attempt to connect to separate machine learning models to a unified framework, and to prove it will do better than individuals through empirical study. However, this paper only shows such combination can achieve better performance than individuals, for the task of segmenting FAQs. Whether there is still an advantage on some other text analysis task, is not clear. It will be even nicer, if theoretical bounds on performance achievable by the combined approach, compared with the bound by either HMM or Maxent individual, can be derived.