Arabic Question Classification using Deep Learning

Abstract

Question classification is one of the essential tasks in Natural Language Processing (NLP). Not many studies have addressed the question classification task where the language is Arabic. This study proposes a language-specific model structure for assessment question classification in Arabic. Approaches proposed in the literature that address Arabic assessment question classification typically have used traditional machine learning techniques. Therefore, the model performance is constrained by the inherent limitations of these traditional approaches. This study addresses this gap and proposes a language-specific feature representation combined with a deep learning model structure. The proposed model uses a modified term frequency-inverse document frequency (M-TF-IDF) term weighting combined with Arabic word embeddings using ArELMo for the question feature representations. The proposed model combines Bi-Directional Gated Recurrent Units (BiGRU) with an attention mechanism and a convolution neural network. The obtained experimental results show the efficacy of the proposed approach which outperforms previous studies on Arabic assessment question classification.