THE GEOMETRY OF SOUND SYMBOLISM

In this multidisciplinary paper, we review research on sound symbolism and present a framework that bridges computational linguistics, machine learning and geomorphometry. The question we pose is whether physical properties of three-dimensional (3D) environments can be used to identify visual meanings associated with English phonesthemes. Phonesthemes are recurring sound sequences that appear in groups of words with related meanings. Drawing on studies such as Godoy et al. (2020), Kawahara et al. (2020) and Alper & Averbuch-Elor (2023), we review the development and linguistic status of phonesthemes. Our study then focuses on the following units: (1) /fl/ for flatness, as in flat and floor; (2) /skr/ for irregularity, as in scraggy and scrape; (3) /sl/ for smoothness, as in slick and slide; and (4) /sp/ for sharpness, as in spike and spear. We created 34 three-dimensional terrains representing these categories and extracted 41 numerical features describing properties such as elevation, slope, roughness, surface orientation and angular variation. These measurements were used to train a multilayer-perceptron neural network in Python. The model achieved 97.06% accuracy in leave-one-out cross-validation. It was then applied, without retraining, to 136 novel terrains and correctly classified 135 of them, producing an accuracy of 99.26%. These findings indicate that the visual categories associated with the selected phonesthemes were consistently represented by multidimensional configurations of physical properties. Our work therefore constitutes an initial step toward linguistic experiments with human participants, who may, in subsequent studies, evaluate nonce words containing the selected phonesthemes in relation to geometrically validated scenes.