Distributional Semantic Model Based on Convolutional Neural Network for Arabic Textual Similarity

2. Problem Statement

The amount of textual information available and stored electronically has grown at a staggering rate. This has exponentially increased the potential source of paraphrase. More formally, given two sentences and , such that , when and convey the same meaning and are semantically equivalent, they are said to be paraphrased (Agarwal et al. 2017). Many researches on paraphrase detection have focused on the English language, but little effort has been done recently on other languages like Arabic. It is considered as a complex problem because of the challenging features of this language (Mohamed et al 2015). It is Semitic spoken by more than 330 million people and composed of 28 letters written from right to left. In addition, Arabic script has a rich morphologically accentuating by the existence of dots, diacritics and stacked letters (Hkiri et al. 2017, Mansouri et al. 2018, Mahmoud et al. 2018). It is highly inflectional, derivational and non-concatenative compared to other languages (Batita et al. 2018, Mahmoud et al. 2017). To contribute and solve these gaps, recent research has been advancing to propose semantic-similarity-based approaches that have more flexibility and expressiveness compared to syntactic ones. The main objective was to measure the degree of relationship between textual units and cover the maximum of Arabic specificities in terms of word construction and diversity meanings.