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

The problem addressed is to develop a model that can reliably identify whether a previously unseen document pair is paraphrased or not. Its detection in Arabic documents is a challenge because of its variability in features and the lack of publicly available corpora. Faced with these problems, the authors propose a semantic approach. At the feature extraction level, the authors use global vectors representation combining global co-occurrence counting and a contextual skip gram model. At the paraphrase identification level, the authors apply a convolutional neural network model to learn more contextual and semantic information between documents. For experiments, the authors use Open Source Arabic Corpora as a source corpus. Then the authors collect different datasets to create a vocabulary model. For the paraphrased corpus construction, the authors replace each word from the source corpus by its most similar one which has the same grammatical class applying the word2vec algorithm and the part-of-speech annotation. Experiments show that the model achieves promising results in terms of precision and recall compared to existing approaches in the literature.