LLM 기반 텍스트 시뮬레이션에서 AI 캐릭터 언어 스타일의 차별성 및 일관성 정량 분석: Big Five 모델을 중심으로

초록

As Large Language Models (LLMs) increasingly drive AI characters, verifying that they maintain their intended personalities remains a challenge that relies heavily on subjective testing. To address this, we propose a quantitative framework utilizing the Big Five (OCEAN) model to objectively measure and validate AI personas. We engineered four distinct GPT-4 Turbo-based characters and analyzed their responses to ten standardized interview questions, focusing on lexical diversity (TTR) and semantic consistency. Results confirmed robust character distinctiveness with a maximum similarity of only 0.28 and a high consistency rate of 96%. Notably, a strong correlation (0.71) between Openness () and TTR reveals a significant engineering link between abstract traits and linguistic output. This framework provides developers with a high-efficiency methodology for systematically tuning and evaluating AI NPC personalities using minimal interaction data.

키워드