Verbal communication and human language possess a very complex and powerful nature in establishing connections and conveying meanings. Poetry is one of the artistic forms of language that, through the combination and arrangement of words and linguistic structures, can beautifully and artistically express emotions, thoughts, and images. This verbal art holds a special place in the Persian language. Great poets, who are the pride of Persian-speaking countries, have elevated and evolved poetry in forms such as Ghazal and Gasidah. On the other hand, a branch of artificial intelligence known as generative AI, alongside natural language processing, has managed to develop models for producing human-like and creative texts that are sometimes used to generate poetry. In this research, we aimed to develop a Persian poetry generation model using attention mechanisms in the Transformer architecture and inspired by the GPT-2 model. This model is capable of managing poetic forms like Ghazal and Masnavi and can imitate the style and pen of several classical poets. For this study, we collected, cleaned, and processed a relatively large training dataset divided into two categories: raw text data and Persian poetry. Additionally, we defined a custom subword tokenizer optimized for the training data. We customized a version of GPT-2 based on our needs. Our Persian poetry generation language model, with approximately 93 million parameters, underwent over 190 hours of training in three phases. Using a custom tokenization strategy designed to understand poetic forms and the poet's style, it achieved the capability of creating Persian poetry while managing poetic forms and styles. The results obtained in this research were analyzed using statistical methods such as BERTScore and BLEU metrics, as well as human eva‎luations, which indicate the model's success in achieving the defined goals of the research

مازيار پالهنگ

مهران صفاياني , سمانه حسيني