Learning of Iconicity in English Words with Neural Language Models

Abstract

This paper discusses experiments designed to answer the question, "How do neural language models learn iconicity in natural language?" We implemented a regression task to predict iconicity ratings of English words using a model based on English BERT. Despite the model not being explicitly informed about the word classes (e.g., noun, verb, onomatopoeia), it demonstrated a tendency for the predicted values to vary according to word class. Identifying the factors contributing to these findings and extending the experimentation to models in other languages represent essential future work.