EmoStance Improves Empathetic AI Responses with Emoji Supervision.

Generating empathetic responses in AI models requires not only understanding what to say but also how to convey the appropriate emotional stance towards the previous speaker's situation. This research formulates this challenge as "response-side affective-orientation control." The EmoStance framework addresses this by leveraging multi-annotator emoji distributions as a form of weak supervision, rather than direct labels, to induce a latent control space that approximates a listener's stance.To facilitate this, the researchers created EmojiDialogue, an extension of the EmpatheticDialogues dataset, enriched with emoji votes and confidence scores at the utterance level. EmoStance then models the emotional expression from the source, predicts a soft response-side orientation based on dialogue context and speaker roles, and uses continuous prefix embeddings to guide a frozen instruction-tuned large language model.Blind pairwise evaluations involving 20 annotators and 800 judgments showed EmoStance achieving a 62.2% decisive win rate. The most notable improvements were observed in the contextual specificity and perceived responsiveness of the generated empathetic replies. This approach also proved complementary to existing external-knowledge methods, offering a novel way to enhance the emotional intelligence of conversational AI.