A Comparative Study of Ensemble Learning Methods for Turkish Text Categorization

Text categorization (TC) is a natural language processing (NLP) task commonly used in text mining. It is defined as the automatic assignment of text to a set of given categories. Ensemble-based methods that unify multiple models effectively improve the prediction performance and result in higher accuracies than a standalone classifier in general. Motivated by this idea, it is proposed to evaluate the performances of the different ensemble classifiers on Turkish TC dataset (i.e., TTC-3600) in this paper. Firstly, OneR attribute evaluator algorithm diminishes the high dimensionality of Turkish text training data. Different combinations of the base classification algorithms are then compared using this reduced data to assess the impact of ensemble learning on Turkish text data. Bayesian network (BN), extra tree (ET), random forest (RF), and multinomial naive bayes (MNB) classification algorithms are used for the base classifiers. bagging, adaBoost, and voting mechanisms are employed to unify these base classifiers in ensemble learning. According to the experimental results, the proposed model achieves a competitive classification accuracy ($93.83 \%$) to the state-of-the-art (SOTA) methods. Moreover, the proposed method has a pretty low runtime that is only 125.8 seconds for TC task.