Overcoming the challenge for text classification in the open world

Classification is often referred to as the task of discriminating one class from others in a given set of classes. Traditionally, classifiers work well assuming that a priori knowledge of all classes are given. Unfortunately, a presenting of unknown class during testing can lead to poor performance of even state-of-the-art classifiers due to observed classes being incorrectly identified to other classes. Recent proposed open world recognition framework provides a promising venue for tackling this challenge. While the majority of works in this relative new field is in computer vision, the rare work in Natural Language Processing shows its instability in its performance and is not based on the open world recognition framework. To tackle this problem, we represent our Nearest Centroid Class (NCC) which is incremental learning and able to detect unknown class during testing. Our model yields promising results in a document classification on text classification domains among current state-of-the-art models.