Deep Neural Network Models for Paraphrased Text Classification in the Arabic Language

Paraphrase is the act of reusing original texts without proper citation of the source. Different obfuscation operations can be employed such as addition/deletion of words, synonym substitutions, lexical changes, active to passive switching, etc. This phenomenon dramatically increased because of the progressive advancement of the web and the automatic text editing tools. Recently, deep leaning methods have gained competitive results than traditional methods for Natural Language Processing (NLP). In this context, we consider the problem of Arabic paraphrase detection. We present different deep neural networks like Convolutional Neural Network (CNN) and Long Short Term Memory (LSTM). Our aim is to study the effective of each one in extracting the proper features of sentences without the knowledge of semantic and syntactic structure of Arabic language. For the experiments, we propose an automatic corpus construction seeing the lack of Arabic resources publicly available. Evaluations reveal that LSTM model achieved the higher rate of semantic similarity and outperformed significantly other state-of-the-art methods.