An automated new approach in fast text classification (fastText): A case study for Turkish text classification without pre-processing

Abstract

Any Text Classification (TC) problem need pre-processing steps which may affect the classification accuracy. Especially pre-processing steps need substantial effort particularly in agglutinative languages such as Turkish. In this context, a traditional text categorization problem requires pre-processing steps such as tokenization, stop-word removal, lower-case conversion, stemming and feature dimension reduction. Before classification, one or more of these steps are applied to text and then a classifier is trained to evaluate the corresponding precision. Deep neural network classifiers combined with word embedding is one of the solutions to eliminate the pre-processing prerequisites. Another novel approach is fastText word embedding based classifier which was developed by Facebook. In this study, we evaluate a fastText classifier on TTC-3600 Turkish dataset without using any pre-processing steps and present the performance of the algorithm.