Unsupervised Learning Enables Extraction of Tactile Information from Text Database

In this work, we propose a new approach to tactile research using natural language processing of archival word corpus as the database. Tactile perception, or assessment of surfaces, is recognized as a language. Thus, by extracting touch-related words and sentences from a text corpus and learning their relationships, we can ultimately learn how humans perceive surfaces. We selected 6 adjectives and 42 onomatopoeias in Japanese as our tactile words. The adjectives represent physical properties, such as roughness and hardness, while onomatopoeias, such as “zara-zara” and “tsuru-tsuru,” are widely used to describe surfaces in Japanese and can correspond to both physical texture cognition and affective cognition. First, using natural language processing of word corpora, we successfully mapped the onomatopoeias with respect to the 6 adjectives, which matched well with the results based on an enquete-based survey. This verified the effectiveness of natural language processing for tactile research. In addition, principal component analysis revealed new tactile dimensions based on onomatopoeias, which we presumably assessaffective tactile dimensions. The proposed approach using natural language processing of archival text databases can provide a large number of datasets for tactile research and culminate in new findings and insights.