Poet Attribution for Urdu: Finding Optimal Configuration for Short Text

Abstract

This study presents a machine learning system to identify the poet of a given poetic piece consisting of 2 lines (i.e. a couplet) or more. The task is more difficult than the general task of author attribution, as the number of words in verses and poems are usually less than the number of articles present in author attribution datasets. We applied classification algorithms with different sets of feature configurations to run several experiments and found that the system performs best when support vector machine using a combination of unigram and bigram are used . The best system (for 5 Urdu poets) has the accuracy of 88.7%.