Automated Analysis of Bangla Poetry for Classification and Poet Identification

Computational analysis of poetry is a challenging and interesting task in NLP. Human expertise on stylistics and aesthetics of poetry is generally expensive and scarce. In this work, we delve into the data to automatically extract stylistic and linguistic information which are useful for analysis and comparison of poems. We make use of semantic (word) features to perform subject-based classification of Bangla poems, and various stylistic as well as semantic features for poet identification. We have used a Multiclass SVM classifier to classify Tagore’s collection of poetry into four categories: devotional, love, nature and nationalism. We identified the most useful word features for each category of poems. The overall accuracy of the classifier was 56.8%, and the analysis led us to conclude that for poetry classification, word features alone do not suffice, due to allusions often being used as a poetic device. We, next, used these features along with stylistic features (syntactic, orthographic and phonemic), for poet identification on a dataset of poems from four poets and achieved a performance of 92.3% using a Multiclass SVM classifier. While contentbased and stylometric analysis of prose in Bangla has been done in the past, this is a first such attempt for poetry.