Directing the flow of conversation in task-oriented dialogue

Next‑generation spoken dialogue systems must support complex goals and multiple parallel tasks, requiring shared control of conversational flow between user and system. This thesis develops a model of conversational initiative grounded in conventions observed in human–human dialogue. Using the DialogueView annotation framework, empirical studies of the Trains and MTD corpora examined how speakers manage subgoals, task switching, and initiative conflicts. Results show that initiative is driven by discourse goals, task switches are signaled to minimize disruption, and conflicts are efficiently resolved using simple cues such as pitch or volume. These findings inform when systems should take initiative, support task switching, and repair conversational breakdowns.