Recurrent multiword units as networks: sequentiality as basis for linguistic generalizations

Abstract Recurrent multiword units (RMUs) are central to language processing, yet their systematic identification and networked organization remain understudied. This study combines corpus-based and network-analytic methods to examine how RMUs contribute to the emergence of constructional schemas. Drawing on a 185-million-word corpus of Taiwan Mandarin, we pursue two aims. First, we propose a quantitative method for identifying cohesive RMUs based on word predictability in context. Second, we model RMUs as a network in which nodes represent RMUs and edges encode structural and semantic similarity, estimated with a state-of-the-art large language model. A comparison with a random sequence network confirms the non-random structure of the RMU network. Analysis of its topology reveals exemplar-based semantic groupings that support higher-level generalizations. These findings highlight RMUs as key building blocks in linguistic categorization, where subgroupings emerge through sequential lexical associations that underlie the formation of grammatical patterns and hierarchical structure.