Hybrid Attention-based Approach for Arabic Paraphrase Detection

Abstract

The growth of data across the web and the ambiguous structure of Arabic language have favored the act of paraphrase. It is defined as a restatement of the original text, giving the same meaning in another form without mentioning its source. Its detection requires calculating semantic textual similarity, which is an important research area in Natural Language Processing (NLP) tasks. Following the literature, deep neural network models have gained satisfactory results in sentence modeling and similarity computing. In this context, a hybrid Siamese neural network architecture is proposed that is composed of the following main components: First, salient features are extracted by applying Global Vectors Representation (GloVe). Then, Convolutional Neural Networks (CNN) capture and learn the contextual meaning of words due to their outstanding performance that has been achieved in different NLP tasks. Then, the output of CNN is combined with an attention model to distinguish the most important words representing the meaning of the sentence. The similarity score between sentences was subsequently computed by applying the cosine measure. Experiments were carried out on a proposed Arabic paraphrased corpus using the Open-Source Arabic Corpora (OSAC). To validate its quality, the SemEval benchmark is used.