Dialogue Systems Benefit from Continuous Addressee Detection

This research re-examines how dialogue systems identify who an utterance is addressed to in multi-party conversations. Traditionally, addressee detection has been treated as a discrete classification task, assigning a single label to an individual or group. However, this paper challenges that assumption, proposing that address is better understood as a continuous phenomenon. Using a multi-party human dialogue corpus with multiple annotator judgments, the study constructed both binary address labels and continuous address levels. The findings indicate that continuous address levels provide a better predictive fit for turn-taking and listener behaviors, including gaze and backchannels, compared to discrete labels. This suggests that the graded structure of address is more accurately captured by continuous representations, opening new avenues for research in more sophisticated and natural dialogue systems.