Arabic Poetry Authorship Attribution and Verification Using Transfer Learning

Abstract

This paper employs a transfer learning approach to attribute and validate the authorship of Arabic poetry. Poetry authorship attribution is a technique of identifying a poet from an anonymous poem related to certain traits. Whereas, poetry verification entails determining whether a poem was written by a specific poet. There are a few works in Arabic poetry recognition, but their fundamental flaw is that they rely primarily on manual feature extraction. This is due to the fact that authors used traditional machine learning approaches. Transfer learning, as part of deep learning, has the ability to extract features from text automatically. Furthermore, using transfer learning to improve text classification accuracy in English has lately shown encouraging results. We employed Arabic-BERT as a transfer learning method for the Arabic language in our research. Arabic-BERT's four models were employed for authorship attribution and authorship verification. In terms of authorship attribution, we found that the best Arabic-BERT model is the large model, which has an f-score of 85%. It outperforms k-nearest neighbours, Naive Bayes, and Support Vector Machines, which are traditional machine learning approaches. In addition, we achieved a score of 96% in authorship verification, outperforming traditional machine learning techniques.