one of the major priblems in question answering system is how to classify a question into a particular class that further will be used to find exact answers within a large collection of documents. kernel dimensionality reduction (KDR) is an alternative method that can be used for features reduction, and in the same time classify question type by using the most effective m-dimensional features in its vector space. in this expriment we used question-answer pairs data from public health domain and word (unigram) features construction. this research shows that KDR correct rate performance is better than SVMafter a head-to-head comparison from 100 observations.