KeisukeKamataki writeup of McCallum 2000

This is a review of Frietag_2000_Maximum_Entropy_Markov_Models_for_Information_Extraction_and_Segmentation by user:KeisukeKamataki.

---

• Summary: They combined MaxEnt and HMM in order to improve original HMM. Combining MaxEnt enabled to handle arbitrary overlapping features and it achieved better from original (Token or Feature) HMM and original MaxEnt. Feature HMM worked (surprisingly) well for the IE task.

---

• I like: I feel this is a very good paper. They are very clear about the problem statement, their approach, experiment and analysis (including reference to confidence interval). The approach of combining MaxEnt sounds reasonable for IE since the task of IE often includes overlapping features such as "word, capitalization, formatting and POS" as written in paper. One of the characteristic properties of MaxEnt to have fewest assumption about the probability distribution of each feature makes easy to plug-in the algorithm to HMM. I was not familiar with GIS, but paper also gives a good explanation about the method.