earticle

영어

Effective emotional support conversations demand nuanced, multi-turn interactions that adaptively employ context-sensitive strategies—an area where large language models (LLMs) often fall short despite their strong general capabilities. To address this gap, we propose a multi-task learning framework that jointly fine-tunes a lightweight DialoGPT model to generate supportive responses and predict the support strategy stage. Using uncertainty-based loss weighting, our method dynamically adjusts multi-task learning objectives based on task-specific uncertainty, enabling balanced optimization between generation and classification tasks. Experiments on the psychologically grounded ESConv dataset show significant improvements, achieving an accuracy of 86.4% and a weighted F1 score of 0.86 in the next-stage strategy prediction task, with particularly strong performance in early dialogue phases such as Exploration. Our study demonstrates that compact LLMs, when guided by task-specific supervision, can effectively deliver strategy-aware emotional support, advancing scalable and reliable mental health conversational agents.