Abstract
Automatic Poem generation is an interesting research topic which has attracted the Natural Language Processing (NLP) deep learning community's attention. However, the majority of poetry-generation-related research is performed in English and Chinese, ignoring other languages. In this paper, we take a first step in training a deep learning model of Arabic poetry generation. We proposed two approaches to generate Arabic poetry: (1) A Bi-directional Gated Recurrent Unit (Bi-GRU) model to compose the first line; (2) A modified Bi-GRU encoder-decoder with a hierarchical neural attention framework to generate other verses sequentially, which can adequately capture word, phrase, and verse information between contexts. A comprehensive evaluation with human judgments confirms that the generated poems by our model outperforms the base models in criteria such as Meaning, Coherence, and Poticness. Extensive quantitative experiments using BLEU scores also demonstrate significant improvements over strong baselines. To the best of the authors' knowledge, this work in generating Arabic poetry is considered an essential development in this line of the field.