ScholarWorks@성균관대학교

초록

Emotional Support Conversation (ESC) systems provide strategic responses to users experiencing negative emotional states to alleviate psychological distress. However, existing embedding similarity-based retrieval methods fail to effectively capture users' implicit emotional needs and subtle contextual cues. Furthermore, the disconnection between retrieval and generation components, coupled with the absence of self-evaluation mechanisms, prevents reliable reflection of strategic intent. This study extends the self-retrieval paradigm-which integrates indexing, retrieval, and self-evaluation within a single large language model (LLM)-specifically for ESC and proposes SAGE, which combines generative retrieval with knowledge fusion. We internalize eight strategies from ESConv as utterance-response-strategy triplets and preserve strategic consistency and emotional nuances through emotion state-based strategy prediction and strategy-specific trie-constrained decoding. We also ensure psychological validity through re-ranking that multiplicatively combines self-evaluation signals with COMET cognitive alignment and HEAL therapeutic appropriateness. In the final generation, we enhance contextual coherence and actionable advice by fusing dialogue context, COMET, and HEAL via cross-attention. On ESConv, strategy accuracy improved by 13.4% over existing LLMs, with consistent superiority in Distinct-1/2, BLEU-4, ROUGE-L, and ESC-Eval/human evaluation. This study integrates strategy-aware internalization, strategy-specific constrained generation, psychological re-ranking, and multi-knowledge fusion generation into a unified end-to-end framework, providing empirical evidence for simultaneous improvements in semantic alignment, strategic balance, and real-world applicability in emotional support conversations.

키워드