Optimizing Threshold Parameters in NLP-based Classification of Koshur Devnagari Scripted Poems

Classification of poetry using Natural Language Processing (NLP) involves categorizing poems into different genres, styles, or themes based on analyzing their linguistic features. The authors, in their previous research paper, had designed a baseline SaSa algorithm for the categorization of Koshur poems into five different categories using the value of threshold as two and received an accuracy of $\mathbf{8 9 \%}$. The threshold value of two means that the tokens with frequencies less than or equal to two will not be considered. It was advantageous because it reduces the algorithm’s time and enhances accuracy. In this study, the authors investigate the effect of threshold variation on the accuracy of a poetry classification model. The experiments were conducted on a diverse dataset of poems to demonstrate how minor adjustments to the threshold value can lead to notable changes in classification accuracy and metrics values. The authors here experimented with threshold values starting from 1 to 5. The findings highlight the importance of finetuning threshold parameters in NLP-based classification tasks for optimal performance.