While large data dependent techniques have made advances in between-genre classification, the identification of subtypes within a genre has largely been overlooked. In this paper, we approach automatic classification of within-genre Bundeli folk music into its subgenres; Gaari, Rai and Phag. Bundeli, which is a dominant dialect spoken in a large belt of Ut-tar Pradesh and Madhya Pradesh has a rich resource of folk songs and an attendant folk tradition. First, we successfully demonstrate that a set of common stopwords in Bundeli can be used to perform broad genre classification between standard Bundeli text (newspaper corpus) and lyrics. We then establish the problem of structural and lexical similarity in within-genre classification using n-grams. Finally, we classify the lyrics data into the three genres using popular machine-learning classifiers: Support Vector Machine (SVM) and kNN classifiers achieving 91.3% and 85% and accuracy respectively. We also use a Naive Bayes classifier which returns an accuracy of 75%. Our results underscore the need to extend popular classification techniques to sparse and small corpora, so as to perform hitherto neglected within genre classification and also exhibit that well known classifiers can also be employed in classifying 'small' data.
展开▼
