Enhancing Artificial Intelligence Empathetic Conversation Generation through Emotion Learning Models

다국어 초록 (Multilingual Abstract)

This study advances empathetic dialogue generation in Korean by integrating emotional learning with artificial intelligence (AI). It addresses the limited understanding of how AI-generated expressions preserve human emotional meaning. A hybrid system was developed using a fine-tuned XLM-RoBERTa model for 24-category emotion classification and ChatGPT-5 for dialogue generation, operating in context-only and emotion-aware modes. Using 18 Korean empathetic dialogue scenarios (108 utterances) from AI Hub, evaluations included exploratory data analysis, BLEU, GLEU, BERTScore, and large language model (LLM)-asa- judge assessments with five external models (ClovaX, Gemini, Perplexity, Claude, and Copilot). Emotion-aware responses were longer (133.7 ± 24.8 characters), more lexically diverse (53.7 ± 8.1 tokens), and preferred by LLM judges in 72.2% of cases, despite comparable semantic similarity (BERTScore > 0.85). The findings highlight the promise of emotion-aware AI for empathetic applications in mental health, education, and customer service, while emphasizing ethical challenges in human–AI interaction.