2023 Volume 19 Issue 3 Pages 105-118
Traditionally, the investigation of the “poetic style” inherent in a collection of Waka was centered around the study of vocabulary distribution and bias. However, with the recent advancements in machine learning, research utilizing word embedding vectors such as word2vec has become feasible. In this study, we demonstrated that by applying principal component analysis to sentence embedding vectors generated by large language models, we can describe the semantic system at the core of Waka's poetic style. Moreover, we observed substantial differences between the “Kokin Wakashū” and “Man'yōshū” and discussed the potential influence of Chinese poetry on these variations.
