Implicit Consequentiality and Topicality in Korean: Evaluating Multiple Discourse-Cue Integration in a Large Language Model

초록

The Linguistic Association of Korea Journal, 34(2), 37-60. This study aims to investigate whether GPT-5.4, one of the latest Large language models (LLMs), shows discourse competence by patterning like native speakers in integrating multiple discourse cues under Korean implicit consequentiality contexts. Adopting a sentence continuation task by Kim and Chun (2023), we tested two cues: verb-based implicit consequentiality (NP1 verbs vs. NP2 verbs) and topicality marking (topic marker vs. nominative marker). Overall, GPT-5.4 exhibited a coreferential bias like native speakers of Korean, preferring subject continuations for subject-referring NP1 verbs and preferring object continuations for object-referring NP2 verbs. However, unlike native speaker patterns, the model showed a significant topicality effect only in NP2 contexts. For referential form selection, we observed that overall the GPT model showed similar preference patterns as native speakers by favoring null subject forms under subject/topic continuity. However, the model used less null forms in object-referring NP2 contexts unlike native-speaker patterns. The results suggest that GPT-5.4 captures overall discourse patterns, but limited in integrating verb semantics, topicality, and discourse accessibility.

키워드