OKAZ: A Deep-learning-Based System for Automatic Arabic Poem Generation

A poem is a sequence of verses that conveys emotions and views via an in-depth knowledge of a language. It represents a deeply rooted element of any language's cultural heritage. For decades, poem generation has been limited to human experts with considerable talent and linguistic skills. However, researchers have been attracted to this area lately, investigating the possibilities of auto-generating poems using AI-based systems. They observe considerable success with highly-inflected languages like Chinese. Nevertheless, research on the auto-generation of Arabic poems is sparse. This work presents one of the first attempts to address this research problem for Arabic poetry by exploring the effectiveness of using deep learning-based models to generate Arabic verses. Our findings show that GPT-2 was able to achieve the best results by outperforming other models and attaining a BLEU score of 0.082 and an average performance of 2.94 out of 5 when evaluated by human experts. We believe that these findings reveal the need for further investigations to bridge the performance gap when compared to the average human-level performance score of 4.62.