An Analysis of Sound Symbolism by Voiced-Voiceless Contrast in Japanese Onomatopoeia Using Word Embeddings

Abstract

It is known that Japanese onomatopoeia has an aspect that has a semantic contrast between voiced and voiceless consonants (e.g., giragira-kirakira) at the beginning of the word. In this paper, we analyzed the difference between voiced and voiceless consonants of onomatopoeia in a space of static word embeddings such as Word2Vec. In an experiment in which we classified onomatopoeia starting with voiced or voiceless consonants in the space of word embeddings, we got a classification accuracy of 0.84 at best, which showed the possibility that word embeddings have some information about the sound symbolism of voiced and voiceless consonants. In an experiment in which we compared the contrast of voiced-voiceless consonants and bipolar adjective pairs, we got a result that showed that particular adjective pairs (e.g., "beautiful-ugly", "bright-dark") have some correlations with the contrast of the consonants, and the correlations have some compatibility with results shown by previous research.