Quality estimation for Japanese Haiku poems using Neural Network

Shinji Kikuchi, Keizo Kato, Junya Saito, Seiji Okura, Kentaro Murase, Takaya Yamamoto, Akira Nakagawa

Abstract: We propose a method to estimate the artistic quality of Haiku (Japanese style short poem) texts using a machine learning approach. Based on the assumption that the artistry of a text stems from its sound factors as well as its meanings, we first constructed two types of vector models, a word-based model and a syllable-based model, converted from Haiku texts. Next, we conducted machine learning for these two models using a convolutional neural network to construct a Haiku quality estimation function. We then evaluated the precision of quality estimation for 40,000 Japanese Haiku poems obtained from a Haiku community site, assuming that the number of “likes” given from viewers to a Haiku corresponds to its artistic quality. Through the experiment, we confirmed that by conducting a quality estimation based on the consensus between different models, we can improve the precision of quality estimation up to 0.64. We also found that if we evaluate Haiku poems for which we have high confidence in quality estimation certainty, the F-measure of the estimation improved from 0.57 to 0.64.