A Morphological Analysis of Phonological Switch Points in Large Language Models-Generated English Blends

Abstract

Forming neologisms such as brunch and smog represents a complex example of linguistic creativity. This process involves the compression of two conceptual domains as well as the intentional manipulation of phonological, orthographic, and semantic material to produce a novel yet interpretable lexical item. With the emergence of Large Language Models (LLMs) as powerful generators of text capable of producing language often indistinguishable from that of humans, the question of whether these models can truly replicate this type of human creativity has become a central focus of inquiry. This study explores the blending capabilities of six contemporary LLMs by implementing a systematic analysis of their neologisms in comparison to those created by humans. Depending on established frameworks of word formation (WF) theory, the study investigates the phonological switch point, the juncture at which the two source words (SWs) are joined to form a new lexeme, in blends generated by LLMs. Through systematic analysis of 43 LLM-generated blends, the study identifies 14 distinct switch point types and reveals that LLMs demonstrate a clear preference for syllable boundary switches, followed by onset-nucleus boundaries and nucleus coda boundaries. The analysis shows that LLMs successfully learn and apply general phonological rules from their training data. These preferences align with those observed in human-generated blends. Furthermore, LLMs are sensitive to phonotactic constraints. They avoid splits within complex consonant clusters. These findings suggest that while LLMs can acquire sophisticated phonological knowledge through statistical learning, their approach to blend formation (BF) reflects learned patterns rather than genuine creative intuition. The study contributes to our understanding of how LLMs process morphological and phonological information, and raises important questions about the nature of linguistic knowledge in Artificial Intelligence (AI) systems.