[4J1-OS-25a] 食とAI(1/2)

[4J1-OS-25a-04] Semantic analysis of complex sizzle words based on ensemble of distributed representation

〇Fumiaki SAITOH¹, Yuki NAGAO¹ (1. Chiba Institute of Technology)

Keywords:Sizzle words, Distributed representation, Customer Review

Sizzle words tends to be widely used in Japanese to easily convey sensory expressions and textures and has been actively studied in recent years. Sizzle words have been used in various ways, and in recent years, new expressions have been created using multiple concepts in a complex manner. For example, a new sense of expression such as "torofuwa" is born from "torotoro" and "fuwafuwa". In this research, word2vec, which can handle semantic word arithmetic, was adopted as a method of acquiring distributed expressions by word embedding, and information on complex concepts is extracted by adding or subtracting words. Furthermore, to avoid the dependency on the initial value of word2vec, this paper also proposes a method to select truly important information using an ensemble of operation results.