ABSTRACT

Tutoring or advising is often performed in spoken face-to-face dialogues. What happens to the structure of user-advisor dialogues when the dialogue is typed and the advisor is a computer system? Typed dialogues in a Wizard-of-Oz setting and spoken face-to-face dialogues are collected and analyzed in three steps: 1) a task structure is derived from an analysis of the users' task ; 2) the task structure is then used to segment the dialogues into hierarchically-related subdialogues producing the dialogue structures ; 3) the distributions of the antecedents of pronominal and of non-pronominal noun phrases in the dialogue structures and the distance between subdialogues in the task structure are compared between the spoken face-to-face and the typed dialogues. The task structure has more influence on the structure of the spoken face-to-face dialogues than on the structure of the typed dialogues: the typed dialogues in the Wizard-of-Oz setting are more similar to a sequence of independent queries than cohesive discourse. This is due to the low interaction level of typed communication and to users' beliefs in poor shared context between users and advisor. Implications for natural language front-ends to advisory systems are generated.