透過回應工程強化生成式人工智慧的客服同理心

Abstract

本研究旨在探討生成式人工智慧（生成式 AI）在顧客服務應用中的回應表現，特別聚焦於同理心、問題解決能力、回應適當性與個人化等面向。為提升生成式 AI 的回應品質，本研究提出結合心理治療概念的回應工程策略，包含角色扮演、示範學習、結合辯證行為療法與現實治療法以及僅採用現實治療法等四種方法，並在不同類型的客戶角色下進行測試。 我們以生成式 AI 模擬顧客與客服間的對話，透過主觀評分（同理心、問題解決與適當性）及語言風格一致性進行評估。結果顯示，所有方法皆能隨迭代次數提升回應品質，而僅採用現實治療法的策略在應對理性顧客時展現優異的問題解決與適當性表現，結合辯證行為療法與現實治療法的策略則更適合處理情緒性顧客的需求。在個人化方面，語言風格一致性表現整體良好，但未呈現穩定上升趨勢，顯示語言風格一致性未必能全面反映回應的個人化程度。 本研究不僅驗證了回應工程策略對生成式 AI 回應品質的增進效益，也提出結合心理療法概念之創新方法，對未來客製化客服 AI 設計與行銷應用具有實務與學術價值。

This study explores how the empathy capabilities of Generative AI (GenAI) can be enhanced for customer service. To improve AI response empathy quality, we proposed four response engineering strategies: (1) Role Play, which instructs the AI to adopt a specific role (e.g., a professional agent) to enhance focus and contextual understanding; (2) Demonstration, which provides the AI with example dialogues to guide its response style and structure through imitation learning; (3) Dialectical Behavior Therapy combined with Reality Therapy (DBT + RT), a hybrid strategy that first guides the AI to acknowledge and regulate emotional expressions (DBT), followed by solution-focused support (RT); and (4) Reality Therapy only (RT-only), which emphasizes present-focused, problem-solving-oriented interactions to address the user's needs directly. These strategies were evaluated across two customer personas—rational and emotional. The experiment involved AI-simulated dialogues between customers and agents, with evaluations conducted through subjective metrics (empathy, problem solving, and appropriateness) and Language Style Matching (LSM) as an indicator of personalization. Results showed that all methods led to improved responses over iterations. Reality therapy only (RT-only)excelled in addressing rational customers, particularly in problem solving and appropriateness, while Dialectical Behavior Therapy combined with Reality Therapy (DBT+RT) proved more effective with emotional customers, especially in empathy-related tasks. However, LSM scores remained high but did not consistently improve, suggesting its limitations in fully capturing response personalization. This research demonstrates the effectiveness of response engineering in enhancing GenAI customer interactions and introduces a novel application of psychotherapy-informed prompting. The findings offer both academic and practical implications for designing personalized AI customer service and marketing systems.