Yandongl writeup of Collins

This is a write up of Discriminative Training Methods for Hidden Markov Models: Theory and Experiments with Perceptron Algorithms, Collins, EMNLP 2002 by user:Yandongl

Traditionally HMM makes use of Viterbi algorithm for training. In this paper authors introduced a new discriminative training method, which is based on the perceptron algorithm by Rosenblatt. These algorithm have been applied to classification tasks by Freund & Schapire and are very effective. Here authors apply it to sequential tagging tasks. New approaches first find the output of the model with Viterbi algorithm, then update the parameters iteratively. Experiment results are promising, but since tasks mentioned in the paper (NP chunking/POS tagging) are well studied, high F-measures are actually not that surprising to me. A comparison between this one and CRF might be better since CRF is considered to be a more modern approach.