Adaptive un-supervised training of translation probabilities using Naïve Bayes word sense disambiguation methods

Abstract

A statistical word sense disambiguation (WSD) model using Naive Bayes assumption is developed in this thesis as a major component for a promising adaptive un-supervised learning algorithm of translation probability. An error reduction of 29.19% to the word sense disambiguation system is reported with the use of the trained translation probability. As the result, a machine translation system applying the word sense disambiguation model together with the trained translation probability of accuracy of 80.16% is achieved. Several variations of the baseline WSD model are implemented and a best variant is identified. Moreover, three alternate WSD models are developed and tested in this work. Experiment results show that the best variant of the baseline model outperforms all the other implementations. Background information and extensive discussions of various information sources are also given in this thesis. Auxiliary data structures has been implemented to allow fast retrieval of the stored information sources.

Date of Award	2000
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology