Paraphrasing Arabic Metaphor with Neural Machine Translation.

The task of recognizing and generating paraphrases is an essential component in many Arabic natural language processing (NLP) applications. A well-established machine translation approach for automatically extracting paraphrases, leverages bilingual corpora to find the equivalent meaning of phrases in a single language, is performed by "pivoting" over a shared translation in another language. Neural machine translation has recently become a viable alternative approach to the more widely-used statistical machine translation. In this paper, we revisit bilingual pivoting in the context of neural machine translation and present a paraphrasing model based mainly on neural networks. Our model describes paraphrases in a continuous space and generates candidate paraphrases for an Arabic source input. Experimental results across datasets confirm that neural paraphrases significantly outperform those obtained with statistical machine translation, in particular the Google translator, and indicate high similarity correlation between our model and human translation, making our model attractive for real-world deployment.