Pattern and Poet Recognition of Arabic Poems Using BiLSTM Networks

Poetry is the oldest form of Arabic literature and continues to play a prominent status in the Arabs' lives. We think that emerging deep learning techniques could provide needed solutions to compose, analyze, process, and enjoy this fine art. We investigate here two related problems: recognizing the poetry pattern of a poetry sample and using machine learning classification techniques to recognize the poet of a poetry verse. Using a dataset of 93,276 diacritized samples and 319 poetry patterns, we demonstrate that deep bidirectional long short-term memory (BiLSTM) network supplemented by a transfer learning solution provides excellent poetry pattern recognition of 99.5%. We also show that the proposed deep BiLSTM network can recognize the poet with 81.2% accuracy when trained on a 100,000-verse dataset of 10 famous Arabic poets. We suggest that the proposed solutions are efficient and need extensions to recognize poetry patterns of undiacritized verses and larger number of poets.