Skip to content
conftrace_
2026 EMNLP EMNLP 2026

Enhancing Emotional Support Conversation with Plug-in Modules for Sequential Strategy Modeling and Strategy-Specialized Generation

AbstractEmotional Support Conversation (ESC) aims to alleviate users’ emotional distress through multi-turn interactions by selecting appropriate support strategies and generating aligned responses. Despite recent progress, existing approaches face three key limitations: (1) the inability to capture sequential transition patterns of strategies within conversations; (2) a training-inference mismatch in strategy prediction; and (3) reliance on a shared decoder that limits strategy-specific expression. To address these challenges, we propose S3-ESC, a simple framework of three components that are model-agnostic across encoder-decoder ESC backbones. First, we introduce a Sequential strategy predictor that models strategy transition patterns. Second, we apply Scheduled sampling for strategy to mitigate the training-inference mismatch. Third, we propose Selective LoRA for strategy that enables strategy-specialized generation. Experiments on six ESC backbones show that S3-ESC consistently improves strategy prediction accuracy, by up to 4.5pp, while enhancing response quality on most of them. Our code is publicly available.

More from EMNLP 2026

All EMNLP 2026 papers →