文献类型：期刊文献

英文题名：Depth vs. Emotional Intensity: Evaluating Empathy inLLMs forEducational Dialog Systems

作者：Mu?oz, Agustín[1]; Schwarzenberg, Pablo[1]; Rojas, Luis[2]; Luo, Fei[3]

机构：[1] Facultad de Ingeniería, Universidad Andres Bello, Santiago, Chile; [2] Facultad de Ingeniería, Universidad San Sebastian, Santiago, Chile; [3] School of Information Science and Engineering, East China University of Science and Technology, Shanghai, 200237, China

年份：2026

卷号：16715 LNCS

起止页码：304

外文期刊名：Lecture Notes in Computer Science

收录：EI(收录号：20263021178654)

语种：英文

外文关键词：Behavioral research - Computational linguistics - Human computer interaction - Speech processing - User interfaces

摘要：This study presents a comparative analysis of two language models of different sizes, Llama3.2:1B-Instruct and Llama3.2:3B-Instruct, within the context of empathetic dialogue. The integration of empathy into dialogue systems is fundamental to improving the User Experience (UX) and fostering engagement and emotional well-being in Human-Computer Interaction. Using the EmpatheticDialogues dataset, 10,943 responses generated by both models to identical prompts were evaluated through a set of automatic metrics designed to capture key aspects of empathetic response quality: lexical specificity (NIDF), emotional modulation (NRC-VAD), emotional intensity (NRC-EIL), empathic dimension labeling (ER, IP, EX labels), and global empathy scores. The findings indicate that Llama3.2:3B-Instruct tends to use a more specific vocabulary (higher NIDF), slightly increases emotional positivity and dominance (valence and dominance), and produces a higher proportion of responses with elevated global empathy. In contrast, Llama3.2:1B-Instruct shows greater emotional variability and more frequent intense emotional reactions (ER_label). However, both architectures exhibit limitations in interpretative (IP_label) and exploratory (EX_label) capabilities, with a strong tendency toward low-depth responses. This bias towards low-depth responses underscores a critical challenge in the design of LLMs for emotional support systems, which must offer richer interpretive and exploratory responses to be effective in some scenarios like education. Overall, while Llama3.2:3B-Instruct demonstrates modest improvements in several aspects, our findings provide empirical guidance for prompt engineers and HCI developers, highlighting the need to implement specific strategies to mitigate the lack of depth and achieve a richer, more coherent empathetic interaction. ? The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.