An improved unsupervised learning probabilistic model of word sense disambiguation

Unsupervised learning can address the general limitation of supervised learning that sense-tagged text is not available for most domains and is expensive to create. However, the existing unsupervised learning probabilistic models are computationally expensive and convergence slowly because of large numbers and random initialization of model parameters. This paper reduces the noise jamming and the dimensionality of the models by using proposed feature selection and initial parameter estimation. Experimental result shows the accuracy and efficiency of the proposed probabilistic model are obviously improved.