Exploring the specificity of linguistic rule learning through reinforcement learning: Semantic and syntactic perspectives

Abstract

Learning linguistic rules is crucial for human cognition, and recent studies have demonstrated that reinforcement learning modeling can effectively simulate rule learning in non-linguistic symbol systems. In this study, we use reinforcement learning to model trial-by-trial dynamic processes of semantic and syntactic rule learning in linguistic symbols (i.e., words in an artificial language) and non-linguistic symbols (i.e., shapes). By analyzing the effects of reinforcement learning parameters on behavioral performance and neural oscillations, we aim to explore whether the mechanisms underlying semantic and syntactic processing differ between linguistic and non-linguistic symbols. Our findings underscore the greater complexity of semantic processing in language, which demands more cognitive resources and engages slower, more deliberative processes. These patterns were reflected by slower response times and a decrease in beta-band power as prediction error signals increased. In contrast, syntactic processing in language—unlike in symbolic tasks—benefited from inherent structural cues, as shown by an increase in beta-band power as prediction error signals grew. These findings provide novel insights into the distinct cognitive and neural mechanisms underlying inherent language rule processing and artificially-created symbolic rule processing within a reinforcement learning paradigm.