KJCKorea
Journal Central

This study aimed to develop an SDGs-based AI creativity and character education program and to examine its educational effects on middle and high school students. The AI Creative Design On (溫)-Namu Project was designed as a TDCD (Thinking–Designing–Creating–Doing) project-based learning model in which students explored social and environmental issues and designed solutions using AI tools. Pre- and post-tests were conducted with students from six secondary schools in the Yeongnam region, and paired-sample t-tests were used for analysis. The results showed significant improvements in students’ interest in AI, creativity, character, and AI problem-solving skills, with large effect sizes. Strong positive correlations among core competencies were also identified, indicating integrated cognitive and affective development. These findings suggest that SDGs-based AI education can serve as a future-oriented educational model that promotes socially responsible and practical problem-solving competencies beyond technical skills.

This study proposes an effective method for building a Text-to-Speech (TTS) model in limited data environments using Korean voice data from elderly speakers aged 60 to 90. We collected approximately 250 minutes of voice data from 50 elderly speakers (25 males and 25 females), applying fine-tuning techniques with the XTTS (Cross-lingual Text-to-Speech) model based on an average of 5 minutes of data per speaker. In the data preprocessing stage, we refined speech segments and transcription quality through Automatic Speech Recognition (ASR) using the Whisper large-v3 model and Voice Activity Detection (VAD). We improved training efficiency and stability by applying Mixed Precision learning and CosineAnnealing scheduler. Experimental results demonstrate that with optimal hyperparameter settings, most speakers achieved low Word Error Rate (WER) and Character Error Rate (CER). Even for speakers who initially showed high error rates during the initial training phase, performance was significantly improved through retraining. This study presents the feasibility of building a Korean TTS system reflecting the voice characteristics of elderly speakers and provides an efficient Few-shot learning methodology applicable in environments with extremely limited data per speaker.

This study aims to develop a conversational learning model designed to support elementary students’ understanding of economic concepts and to empirically validate the educational feasibility of Large Language Models (LLMs). To this end, a custom dataset comprising 1,000 question-and-answer pairs was constructed, capturing everyday economic scenarios closely related to the lived experiences of elementary students. A domain-specific model for economic education was then developed by applying Low-Rank Adaptation (LoRA)-based fine-tuning to the LLaMA 3.1 8B Instruct model. Experimental results demonstrated that the fine-tuned model achieved a substantial performance gain, with the Bilingual Evaluation Understudy (BLEU) score increasing from 0.18 to 5.67—approximately a 31-fold improvement over the base model. Furthermore, the BERTScore (F1), which measures semantic similarity, rose from 0.6642 to 0.7743, confirming enhanced quality and accuracy in sentence generation. These findings suggest that Parameter-Efficient Fine-Tuning (PEFT) enables the effective construction of domain-specific language models even within environments characterized by compact datasets and limited GPU resources. This study establishes an empirical foundation for designing LLM-based economic education systems for primary learners and highlights the potential for evolving into personalized, adaptive educational tools through future user-centered evaluations and dataset expansion.