Scholar Project: Application of Discriminative Training Method with Perceptron Algorithm for Hidden Markov and Conditional Random Field Models
■ 01.2017 - 02.2017 ■ Application of Discriminative Training Method with Perceptron Algorithm for Hidden Markov and Conditional Random Field Models
○ Software Dev (Utility)
○ Project: 2 months in Paris-Saclay University (PSU)
○ Supervisor: Florence d’Alché-Buc {florence.dache@telecom-paristech.fr }
○ Technical Keywords: HMM, Perceptron algorithm, CRF, maximum-likelihood, Viterbi algorithm, python
○ Github
We apply the discriminative training method for Hidden Markov and Conditional Random Field(CRF) models to typos-correction problem in Natural Language Processing. The method, proposed by Michael Collins in 2002, is based on perceptron algorithm, in which Viterbi algorithm is used for decoding the data sample and then the parameters are updated iteratively. We show experimental results where we compare this method with Maximum Likelihood estimation for HMM and CRF model trained with this method compared to CRF trained with Maximum Likelihood.
○ Project: 2 months in Paris-Saclay University (PSU)
○ Supervisor: Florence d’Alché-Buc {florence.dache@telecom-paristech.fr }
○ Technical Keywords: HMM, Perceptron algorithm, CRF, maximum-likelihood, Viterbi algorithm, python
○ Github
We apply the discriminative training method for Hidden Markov and Conditional Random Field(CRF) models to typos-correction problem in Natural Language Processing. The method, proposed by Michael Collins in 2002, is based on perceptron algorithm, in which Viterbi algorithm is used for decoding the data sample and then the parameters are updated iteratively. We show experimental results where we compare this method with Maximum Likelihood estimation for HMM and CRF model trained with this method compared to CRF trained with Maximum Likelihood.