Treniranje strukturnih klasifikatora za različite funkcije gubitaka sa primenom na probleme klasifikovanja sekvenci

Abstract

This thesis presents algorithms for training structured classifiers over different loss functions. It introduces a new primal subgradient method for the optimization of average sum loss, several extensions of max-margin classifiers to the k-best case, a sequential dual method for the structured ramp loss optimization, It considers different decoding algorithms over semirings and presents an organization of a two-structured-model committee in order to improve the results. It includes theoretical analysis of introduced algorithms, as well as experimental results on sequence labelling problems in natural language processing.