A short note on the structural priming in LLM: Focusing on dative constructions in Korean

Recently, with the increasing diversity of research on neural language models, there has been growing discussion about whether the outcomes of purely probabilistic learning models reflect the so-called “linguistic knowledge” assumed in theoretical linguistics. This study investigates whether neural language models acquire structural cues of Korean dative constructions (DACs) so as to distinguish it from superficially similar constructions, in terms of structural priming. Using three large language models (LLMs)―LLaMA 3.1, Bllossom, and Polyglot-Ko ―we calculate and compare the priming effects based on the probabilities, by examining various structural properties of our four target constructions (i.e., DAC, DAC_Source, Causative, and Passive). The results reveal that at least the current LLMs employed in this study are capable of recognizing and distinguishing the abstract structures of our target constructions. Concomitantly, it is also observed that (i) cognitive concepts associated with spatial perception seem to have a stronger influence on structural priming than theta-roles, and (ii) the structural features of certain postposition/dative markers play a crucial role in differentiating the target sentences.