Enhancing the performance of TF-IDF and Doc2Vec based turkish text categorization system with phrase modeling

TF-IDF term weighting measure is based on frequency of words in texts. This measure doesn't capture the semantic relationship between words. Doc2Vec which is based on artificial neural networks can capture the semantic relations between the words and it enables to yield document vectors of a more manageable size. Consecutive word detection has been reported to have important effects on text mining by many studies. Consecutive word phrases are important for expressing the semantic integrity within the texts. In this study, the performances of traditional machine learning classifiers and ensemble learning algorithms are compared on four different Turkish datasets which are vectorized with both traditional TF-IDF term weighting measurement and Doc2Vec method. Naïve Bayes, K-Nearest Neighborhood, Logistic Regression, Support Vector Machines, Decision Trees, Multilayer Perceptrons algorithms are used as base classifiers. Beside these classifiers Random Forest, Bagging and Adaboost algorithms are also used as ensemble learning approaches. The three most successful classification methods are combined with Majority Voting and the classification results are shared. The classifiers have been applied on 4 different Turkish datasets containing news documents of different lengths. The contributions of our study are "to apply consecutive word detection process to the documents before the classification phase" and "to show that the performances of the applied classifiers' results have been increased after the consecutive word detection phase is applied". In addition to the approach based on frequency of words for consecutive word detection, we also use the url links of Turkish Wikipedia. By using consecutive word detection, higher performance values are presented in almost all classification experiments.