Which is Stronger? : Discriminative Learning of Sound Symbolism


Abstract: The importance of sound symbolism of a product name has been emphasized in recent researches. However, detailed mechanisms of sound symbolism remain controversial, even in reports of recent studies. This study examines a method to detect sound symbolism using discriminative learning. First, we build a training dataset comprising name pairs: (1) name-A, (2) name-B, and (3) a label showing which has stronger sound symbolism. Next, we train a Support Vector Machine (SVM) to learn data using both character-based features and phoneme-based features. In experiments, the proposed method demonstrated almost identical performance to that of humans (72% agreement ratio to humans). This paper also presents a method to generate new names with strong sound symbolism based on greedy search. The generated names have high agreement to human judgments (84%). This the first study suggesting that machine learning can detect sound symbolism.

Back to Table of Contents