Interpretable Predictive Results in Classification of Waka Poets

Abstract

In this paper, we present our proposed method for providing interpretable predictive results in the classification of waka poets. Waka is a type of traditional Japanese poetry. Waka poetry reflects the mind of waka poets. By extracting taste and feeling from waka poetry, it is possible to reveal aesthetic notions in waka poets. This work focuses on waka poetry composed by the Sanjurokkasen included in the Nijuichidaishu. We thus interpret the predicted results in the classification of waka poets by LIME. The number of waka poetry data varies from poet to poet. Therefore, we adopt Naive Bayes as the classification algorithm in this work, which is expected to have high accuracy even when the number of data is small. Using LIME enables interpretation of results in the classification of waka poets; it is possible to provide rationale words for the classification results. The study of waka poetry from the point of view of data science contributes to the development of literary study by providing objective verification of hypotheses and new findings such as previously undiscovered expressions and interpretations.