Open-access Dataset on Acceptability Ratings of Korean Clausal Constructions by Humans and GPT Models

Abstract

The present study introduces a new, open-access dataset on acceptability ratings of Korean clausal constructions at the morphosyntax–semantics interface (dative, passive, and negative polarity item). The dataset comprises (i) linguistically controlled sentence materials, (ii) ratings from targeted adult populations (individuals in their 20s), and (iii) parallel ratings from GPT variants (including ChatGPT). Alongside the release, we assess the alignment between GPT- and human-derived ratings to probe the extent to which GPT architectures can approximate patterns of human sentence comprehension.

Resources

Details