Understanding what is being communicated in a dialogue involves determining how it is coherent, that is, how the successive turns in the dialogue are related, what the speakers’ intentions, goals, beliefs, and expectations are and how they relate to each other’s responses. This thesis aims to address how turns in dialogue are related when one speaker indicates contrast with something in the preceding discourse signalled by “but”. Different relations cued by “but” will be distinguished and characterised when they relate material spanning speaker turns and an implementation in a working dialogue system is specified with the aim of enabling a better model of dialogue understanding and achieving more precise response generation. A large amount of research in discourse addresses coherence in monologue, and much of it focuses on cases in which the coherence relation is explicitly signalled via a cue-phrase or discourse marker (e.g., “on the other hand”, “but”, et cetera) which provides an explicit cue about the nature of the underlying relation linking the two clauses. However despite research on Speech Acts, planning research into speakers’ intentions, and semantic approaches to question-answering dialogues, very little work has focused on coherence relations across turns in dialogue even given the presence of a cue-phrase. This thesis will explore what sorts of relations the speaker of the “but” perceives between elements in the dialogue, and in particular, it will focus on “but”s communicating Denial of Expectation, Concession, and Correction by determining what underlying cross-turn expectations are denied in the former two, and what is being corrected in the latter case. We will extend work by Lagerwerf (1998) in monologue which presents a treatment for Denial of Expectation and Concession arguing that “but” implicates a defeasible expectation which is then denied (in Denial of Expectation) or argued against (in Concession). We also follow Knott’s approach (Knott, 1999a) of describing the semantics of a cue-phrase algorithmically from the agent’s mental model of the related utterances. Task-oriented and nontask-oriented spoken dialogues involving turn-initial “but” are examined, motivating a logical scheme whereby Denial of Expectation, Concession and Correction can be distinguished. These relations are then modelled in the PTT (Poesio and Traum, 1998) Information State (Matheson, Poesio and Traum, 2000) model of dialogue, enabling more relevant response generation in dialogue systems.

As speech and natural language processing technology advance, it now reaches a stage where the dialogue control or initiative can be studied to realise usable and friendly human computer interface programs such as computer dialogue systems. One of the major problems concerning dialogue initiative is who should take the dialogue initiative when. This thesis tackles this dialogue initiative problem using the following approaches: 1. Human dialogue data is examined for their local dialogue structures; 2. A dialogue manager is proposed and implemented, which handles variations of human dialogue data concerning the dialogue initiative, and experimental results are obtained by having the implemented dialogue managers working with a parser and a generator exchange natural language messages with each other; and 3. A mathematical model is constructed and used to analyse who should take the dialogue initiative when. The first study shows that human dialogue data varies concerning the number of utterance units in a turn and utterance types independently of the difference of the dialogue initiative. The second study shows that the dialogues in which the dialogue initiative constantly alters (mixed-initiative dialogues) are not always more efficient than those in which the dialogue initiative does not change (non mixed-initiative dialogues). The third study concludes that under the assumption that both speakers solve a problem under similar situations, mixed-initiative dialogues are more efficient than non-mixed-initiative dialogues when initiating utterances can reduce a problem search space more efficiently than responding utterances. The above conclusion can be simplified to the condition that the agent should take the dialogue initiative when s/he can make an effective utterance like in the situations where s/he has more knowledge than the partner with respect to the current goal.