Using Language Models for Text Classification

This paper describes an approach to text classification using language models. This approach is a natural extension of the traditional Naïve Bayes classifier, in which we replace the Laplace smoothing by some more sophisticated smoothing methods. In this paper, we tested four smoothing methods commonly used in information retrieval. Our experimental results show that using a language model, we are able to obtain better performance than traditional Naïve Bayes classifier. In addition, we also introduce into the existing smoothing methods an additional factor of smoothing scale according to the amount of training data of the class, and this allows us to further improve the classification performance.