A learning approach towards metre-based classification of similar Hindi poems using proposed two-level data transformation.

The article focuses on the development of a machine learning-based system for metre-based classification of Hindi poetry, addressing a gap where prior work predominantly used rule-based methods. It introduces a novel two-level data transformation process—converting Hindi poetic lines into moraic (mātrā) sequences and then generating prefix sums—to extract nineteen engineered features that effectively distinguish six similar moraic metres (Rolā, Soraṭhā, Tomara, Rūpamālā, Śobhana, and Dohā). Using a dataset of 2,922 lines from a prominent Hindi poetry repository, the study applies Bernoulli Naïve Bayes, k-nearest neighbour, random forest, and support vector machine classifiers, achieving up to 99.65% accuracy with random forest. A post-processing step is incorporated to address overlapping class issues between Rūpamālā and Śobhana metres, enhancing classification clarity. This work provides a benchmark dataset and a methodological framework that can be extended to broader Hindi poetic forms, offering a significant contribution to computational analysis of Hindi poetry.