Implicit Causality in Korean: Evaluating GPT-4o’s Discourse Competence

초록

This study evaluates GPT-4o’s ability to reproduce human-like discourse biases in Korean Implicit Causality (IC) contexts, probing both reference resolution and form-function mappings. Building on Kim (2019) and Kim & Grüter (2019), GPT-4o was presented with Korean sentence fragments – object-biased IC (IC2) verbs, subject-biased IC verbs with a syntactic causative (SC) marker (IC1-SC), and subject-biased IC verbs without the syntactic causative marker (IC1-Non-SC) – using In-Context Learning method. GPT-4o’s continuations were compared to native-speaker data for (i) which antecedent was re-mentioned after because, and (ii) the choice of referential form (repeated name, null subject, overt pronoun, full NP). GPT-4o exhibited a robust IC-verb effect, favoring subject continuations for IC1 and object continuations for IC2 verbs, but showed no sensitivity to the SC marker. Its overall form profile aligned with human patterns but diverged significantly in statistical tests. These findings suggest that, while GPT-4o captures broad IC semantics, it remains limited at the morphology-discourse interface, underscoring the need for more balanced multilingual pretraining and fine-tuning on low-resource languages.

키워드