Abstract
Text classification, whether by topic or genre, is an important task that contributes to text extraction, retrieval, summarization and question answering. In this paper we present a new pairwise ensemble approach, which uses pairwise Support Vector Machine (SVM) classifiers as base classifiers and “input-dependent latent variable” method for model combination. This new approach better captures the characteristics of genre classification, including its heterogeneous nature. Our experiments on two multi-genre collections and one topic-based classification datasets show that the pairwise ensemble method outperforms both boosting, which has been demonstrated as a powerful ensemble approach, and Error-Correcting Output Codes (ECOC), which applies pairwise-like classifiers for multiclass classification problems.
