Enhancing Support Vector Machine Performance Using Particle Swarm Optimization for Arabic Text Classification

Abstract

Arabic is a very complex language, stemming from its high morphological richness, high linguistic complexity, and inadequate sizes of available quality labeled resources. In this paper, we attempt to address such obstacles by proposing a novel hybrid model that combines SVM and PSO for improving Arabic text classification. The efficacy of SVMs for classification is known to be good, their performance depends on hyperparameter settings. The data has been classified using SVM with optimized parameters (C, Gamma, Kernel type and feature subset) on the basis of PSO technique which is a strong metaheuristic algorithm. The experimental results of the proposed SVM-PSO were evaluated with a corpus composed from different on line sources in Arabic, these representing multiple lineage dealing with various categories such as news (sports, politics), history, geography and IT. Experiments show that the proposed method not only results in significantly better performance than compared methods with a large advantage, but also high classification accuracy (99.2%). The model was also found to have substantially greater precision (99.1%), recall (99.0%), and F-Score (99.1%) as compared to traditional machine learning classifiers, namely standard SVM (92%), Naive Bayes (88%) and Random Forest (94%). This research showed that combining metaheuristic search for parameter tuning, along with classical machine learning can result robust classifier to process Arabic text as one of those state-of-the-art ones in the fair field of NLP.