Neural Symbolic Arabic Paraphrasing with Automatic Evaluation

摘要

We present symbolic and neural approaches for Arabic paraphrasing that yield highparaphrasing accuracy. This is the first work on sentence level paraphrase generation forArabic and the first using neural models to generate paraphrased sentences for Arabic. Wepresent and compare several methods for para- phrasing and obtaining monolingual paralleldata. We share a large coverage phrase dictionary for Arabic and contribute a large parallelmonolingual corpus that can be used in developing new seq-to-seq models for paraphrasing.This is the first large monolingual corpus of Arabic. We also present first results in Arabicparaphrasing using seq-to-seq neural methods. Additionally, we propose a novel automaticevaluation metric for paraphrasing that correlates highly with human judgement.