KeisukeKamataki writeup of writeup of Sha and Pereira 2003

From Cohen Courses
Jump to navigationJump to search

This is a review of sha_2003_shallow_parsing_with_conditional_random_fields by user:KeisukeKamataki.

  • Summary:

This paper applied CRF with different training methods for shallow parsing (extracting NP chuncks from sentences). They reported the best performances of each training method and also analyzed relation between training speed and its performance of each method. Unlike HMM, CRF tries to model hidden states with global feature vector (and weighting function) of observed global sequence. This approach improved F-measure score from MEMM to a little bit and achieved much better accuracy than MEMM. Except for GIS, which is a baseline method, most of the training methods they test achieved similar F-measure, but learning time was significantly different.

  • Not clear: When they compare performance, they only show the best result. But for practice, average and/or variance performance information might be also important. Couldn't well understand L-BFGS.
Retrieved from "http://curtis.ml.cmu.edu/w/courses/index.php?title=KeisukeKamataki_writeup_of_writeup_of_Sha_and_Pereira_2003&oldid=689"