著者

下窪 聖人

出版者

法政大学大学院情報科学研究科

雑誌

法政大学大学院紀要. 情報科学研究科編 (ISSN:24321192)

巻号頁・発行日

vol.16, pp.1-6, 2021-03-24

---

In recent years, E-book provides many opportunities that people read books during a commute or break time. Though an E-book reader provides some useful functions for reading the books. This system enables a new function of recommending illustrations during one is reading. This is achieved by calculating cosine similarity which uses the BERT based distributed representation of each scene. There are mainly 4 steps: (1) this system divides texts into some scenes by TopicTiling; (2) the important sentences are extracted by using BERT for each scene; (3) subjects, verbs, and objects are extracted using MeCab and CaboCha; and (4) the system recommends appropriate illustrations by calculating similarity between the sentences by using the analysis results and image titles. As expected, this system can recommend illustrations for a reader to form the image when he is difficult to understand E-book in only text content. These recommended illustrations are original from the illustrations of the web site “irasutoya”. As a result of the experiment, the accuracy of the illustration’s recommendation was about 72.0 percent. This result was obtained by using CNN to convert the image style of both the original and the recommended illustrations, and then evaluating them by SSIM.