Earticle

초록

영어

Building on Jang(2021), we test whether text-only LLMs show graded evaluation of written Korean aegyo and how these relate to (un)naturalness language in explanations. We generate 77 informal past-tense verb forms from 11 high-frequency stems: a standard spelling (S), a resyllabified baseline (S′), and alternations from S′ (/j/-insertion, affrication, stopping, and combinations). We query GPT-4o, o3, and GPT-5, collecting 1–7 cuteness ratings and by-item justifications. All models categorically recognize aegyo, rating manipulated forms well above S and matching Jang(2021) by placing CJ at the top. Differences emerge in how strongly they demote mid-tier variants judged less conventional. A cue-based justification analysis reveals that mentions of ‘awkward/unnatural/nonstandard’ predict lower ratings, whereas generic ‘cute/endearing’ language is near-ubiquitous. Written aegyo thus separates register detection from gradient evaluation of (un)naturalness under text-only conditions, making it a useful diagnostic probe and benchmark candidate as opposed to definitive tests of phonotactic competence.

목차

Abstract
1. Introduction
2. Background and research questions
2.1. Aegyo and the linguistic structure of aegyo
2.2. Perception studies: phonological features and phonotactics
2.3. Prior LLM work on aegyo as a benchmark
2.4. Research questions
3. Method
3.1. Models
3.2. Stimuli
3.3. Prompting and response protocol
3.4. Explanation coding and statistical analysis
4. Results
4.1. Model ratings of aegyo and standard forms
4.2. Explanations as evidence of perceived (un)naturalness
5. General discussion
5.1. Endpoint agreement and mid-tier divergence
5.2. From ratings to rationales
5.3. Rating scale and ranking patterns
5.4. Implications and next steps
5.5. Limitations and scope
6. Conclusion
References

키워드

애교 대규모 언어 모델 음운 배열 제약 대규모 언어 모델 평가 관습성 사용 빈도 표기적 두드러짐 aegyo large language models phonotactic constraints LLM evaluation conventionality usage frequency orthographic salience

저자

• Rok Sim [ University of South Carolina/PhD candidate ] First author
• Drew Crosby [ Korea University/Lecturer ] Corresponding Author

참고문헌

발행기관

간행물