Detecting Meter of Arabic Poetry Using Deep Learning: Testing Parameters and Model Design

Arabic poetry adheres to specific rhythmic patterns known as meters. Traditionally, rule-based meter identification is a complex process that involves converting the poem into a special form (Arud form) and then analyzing its components. This paper explores the application of Natural Language Processing (NLP) with pattern recognition for Arabic poetry meter classification using Recurrent Neural Networks (RNNs). We evaluated various RNN architectures, including BiLSTM and BiGRU, to identify the most effective model for this task, considering several factors that may affect classification accuracy, such as overfitting, bias toward the most represented categories in the dataset, and diacritics. Unlike traditional methods that rely on diacritics, our approach classifies meters directly from text, enhancing efficiency and applicability even with partial or no diacritics. Finally, we proposed a BiLSTM model achieving a classification accuracy of up to 96.57%. This model was designed to classify effectively with fewer training examples, thereby saving time and reducing computational resource consumption, making it superior to its peers. We also reviewed relevant deep learning studies to situate our work within the broader meter classification landscape.