Incorporating External Information in Bayesian Classifiers Via Linear Feature Transformations

摘要

Naive Bayes classifier is a frequently used method in various natural language processing tasks. Inspired by a modified version of the method called the flexible Bayes classifier, we explore the use of linear feature transformations together with the Bayesian classifiers, because it provides us an elegant way to endow the classifier with an external information that is relevant to the task. While the flexible Bayes classifier is based on the idea of using kernel density estimation to obtain the class conditional probabilities of continuously valued attributes, we use the linear transformations to smooth the feature frequency counts of discrete valued attributes. We evaluate the method on the context sensitive spelling error correction problem using the Reuters corpus. For this particular task, we define a positional feature transformation and a word feature transformation that take advantage of the positional information of the context words and the part-of-speech information of words, respectively. Our experimental results show that the performance of the Bayesian classifiers in the natural language disambiguation tasks can be improved with the proposed transformations and that the incorporation of external information via the linear feature transformations is a promising research direction.