대규모 언어 모델을 이용한 마비말장애 환자의 발화 분석 및 맞춤형 재활 제안 챗봇 시스템

초록

This study proposes a chatbot system based on a large language model (LLM) to analyze the speech of patients with dysarthria and suggest personalized language rehabilitation plans. Traditional speech evaluation methods heavily rely on clinical experts, incur high time costs, and are prone to subjective judgment. To address these challenges, this study employed GPT-4-turbo-preview to analyze utterance data from 347 dysarthric patients (severity 0: 12, severity 1: 332, severity 2: 3) using the Autumn Paragraph reading task. The system offers three core functions—pronunciation practice assistance, inappropriate pause analysis, and daily conversation—through a chatbot interface that uses special tokens to represent pause intervals. The inappropriate pause analysis feature automatically assesses patient speech patterns, accurately identifying the speech characteristics of dysarthric patients and providing tailored rehabilitation therapies. Evaluation using G-EVAL yielded scores of 3.89 for coherence, 4.85 for consistency, and 4.84 for relevance.

키워드