Effectiveness of zero-shot models in automatic Arabic Poem generation

Abstract

Text generation is one of the most challenging applications in artificial intelligence and natural language processing. In recent years, text generation has gotten much attention thanks to the advances in deep learning and language modeling approaches. However, writing poetry is a challenging activity for humans that necessitates creativity and a high level of linguistic ability. Therefore, automatic poem generation is an important research issue that has piqued the interest of the Natural Language Processing (NLP) community. Several researchers have examined automatic poem generation using deep learning approaches, but little has focused on Arabic poetry. In this work, we exhibit how we utilize various GPT-2 and GPT-3 models to automatically generate Arabic poems. BLEU scores and human evaluation are used to evaluate the results of four GPT-based models. Both BLEU scores and human evaluations indicate that fine-tuned GPT-2 outperforms GPT-3 and fine-tuned GPT-3 models, with GPT-3 model having the lowest value in terms of Poeticness. To the best of the authors' knowledge, this work is the first in literature that employs and fine-tunes GPT-3 to generate Arabic poems.