Improving Multiclass Text Classification with Error-Correcting Output Coding and Sub-class Partitions

Abstract

Error-Correcting Output Coding (ECOC) is a general framework for multiclass text classification with a set of binary classifiers. It can not only help a binary classifier solve multi-class classification problems, but also boost the performance of a multi-class classifier. When building each individual binary classifier in ECOC, multiple classes are randomly grouped into two disjoint groups: positive and negative. However, when training such a binary classifier, sub-class distribution within positive and negative classes is neglected. Utilizing this information is expected to improve a binary classifier. We thus design a simple binary classification strategy via multi-class categorization (2vM) to make use of sub-class partition information, which can lead to better performance over the traditional binary classification. The proposed binary classification strategy is then applied to enhance ECOC. Experiments on document categorization and question classification show its effectiveness.