Phonology meets AI: Can ChatGPT detect and differentiate cuteness in Korean aegyo?
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
This study investigates whether ChatGPT can recognize the affective force of aegyo, a Korean speech style marked by playful phonological alternations such as nasalization, affrication, and /j/-insertion. More specifically, it examines whether ChatGPT detects aegyo forms as “cuter” than their canonical counterparts and whether it can distinguish degrees of cuteness based on phonological features and phonotactic acceptability, following Jang (2021). Three ChatGPT versions (GPT-3.5, GPT-4, GPT-4o) rated 11 verbs in seven conditions on a 7-point scale. Results showed that aegyo forms were rated as cuter, indicating sensitivity to affective cues. However, internal rankings diverged from human judgments: while humans favored forms aligned with phonotactic constraints—especially affrication and stopping—ChatGPT prioritized /j/-insertion. Version-specific trends appeared: GPT-3.5 gave conservative scores, GPT-4 moderate contrasts, and GPT-4o stronger contrasts with greater variability, suggesting partial but limited sociophonetic sensitivity.
Abstract
This study investigates whether ChatGPT can recognize the affective force of aegyo, a Korean speech style marked by playful phonological alternations such as nasalization, affrication, and /j/-insertion. More specifically, it examines whether ChatGPT detects aegyo forms as “cuter” than their canonical counterparts and whether it can distinguish degrees of cuteness based on phonological features and phonotactic acceptability, following Jang (2021). Three ChatGPT versions (GPT-3.5, GPT-4, GPT-4o) rated 11 verbs in seven conditions on a 7-point scale. Results showed that aegyo forms were rated as cuter, indicating sensitivity to affective cues. However, internal rankings diverged from human judgments: while humans favored forms aligned with phonotactic constraints—especially affrication and stopping—ChatGPT prioritized /j/-insertion. Version-specific trends appeared: GPT-3.5 gave conservative scores, GPT-4 moderate contrasts, and GPT-4o stronger contrasts with greater variability, suggesting partial but limited sociophonetic sensitivity.