Recent emotional support dialogue systems mainly focus on one‑to‑one seeker‑supporter interactions, considering only individual emotional states while overlooking the evolving relationships in multi‑party settings.
We therefore introduce relation‑aware emotional support conversation as a new task to test whether large language models (LLMs) can capture and exploit relational dynamics for more effective support. Using real couple and family interview recordings we build RESCUE (Relation-aware Emotional Support Conversation Understanding and Evaluation Benchmark), which contains 191 samples, 7,079 annotated turns, and 1,064.8 minutes of video.
The benchmark provides fine‑grained socio‑emotional and support‑related annotations and defines six evaluation tasks that target relational understanding and relation‑sensitive support, such as relation pattern prediction, viewpoint prediction, and support strategy prediction.
Experiments with ten mainstream LLMs show decent performance on tasks that rely on local emotional or intervention cues, but a marked drop on relation‑intensive tasks, revealing current models’ limitations in modeling interpersonal relations and making relation‑aware support decisions.
Review