Kyoto researchers redefine addressee detection as a continuous spectrum in multi-party AI dialogue

In multi-party dialogue systems—think AI assistants conversing with multiple users—knowing who is being addressed is critical. Traditionally, this is treated as a multi-class classification: pick one label (a specific user or the group). But a new paper from Kyoto University, led by Taiga Mori, argues that address is actually a continuous spectrum. They analyzed a human-human multi-party dialogue corpus where each utterance was annotated by multiple raters. Instead of forcing a single label via majority vote, they used a latent-variable model to infer a continuous 'address level' from the annotator disagreement.

The results are striking: continuous address levels not only predict turn-taking better than discrete labels, but they also correlate strongly with listener behaviors like gaze direction and backchannel responses (e.g., nodding, 'uh-huh'). This suggests that speakers implicitly address others on a gradient—not just 'person A' or 'everyone.' The implication for AI: future dialogue systems could use continuous address representations to achieve more natural group interactions, better manage multi-user contexts, and reduce awkward turn-taking failures. The paper is a methodological shift that could influence how Alexa, Google Assistant, or next-gen smart speakers handle group conversations.