Sound symbolism in pseudoword generation

Abstract

In this study, artificial intelligence (AI) and rule-based systems were used to explore sound symbolism guided pseudoword generation. 4 models were developed and evaluated: 3 AI based, and 1 rule based. The input for each model is a scalar value between 0 (sharp) and 1 (round) of the desired phonetic quality of the generated pseudoword. Although the AI models had difficulty learning sound symbolic relationships effectively, the results suggest that a larger, more reliable dataset may improve model performance. An unexpected outcome from this study was the development of a BERT based model to attempt to predict the roundness of pseudowords. Although the pseudoword generators had limited generation capability and the dataset was limited, this project shows the feasibility of using AI models to create pseudowords that are consistent with sound symbolic relationships.