By proposing to focus on discourse obligations for explaining the actions of speakers in conversation, Traum and Allen (1994) initiated a line of re- search developed further by Poesio and Traum (1997; 1998), Matheson et
9This is similar to intention-based approaches to cooperation in which systems assume
the user’s intentions as their own and then reason to act following those intentions. Galliers (Galliers, 1988) calls this kind of unconditional cooperation benevolence.
al. (2000) and Kreutel and Matheson (1999;2000;2001;2003b). The model of conversational agents presented in Chapter 5 draws elements from these works, which we summarize below.
Poesio and Traum Theory (PTT)
PTT was first proposed by Poesio and Traum (1997;1998) on the basis that participants’ actions in a conversation be part of the common ground, with the aim of providing a unified treatment for discourse context in reference resolution, intention recognition and dialogue management. The authors claim that including explicitly the occurrence of conversation acts, as op- posed to just their (domain-dependent) propositional content, as part of the dialogue situation that is agreed upon between participants, represents a shift from modelling the meaning of contributions to modelling their use. This, in turn, allows considering pragmatic information in the deliberation process for managing the dialogue.
The dialogue acts in PTT are taken from Conversation Act (CA) theory (Traum and Hinkelman, 1992): an extension of the theory of speech acts, to consider aspects from turn-taking, adjacency pairs and grounding. In speech act theory, a locutionary act (i.e. an utterance) usually generates several illocutionary acts (e.g. inform, request, accept). These are called core speech acts in CA and are required to be grounded before taking full effect. Grounding is achieved by means of grounding acts (e.g. acknowledge, request repair). CA also considers turn-taking (e.g. take turn, keep-turn, release-turn) and argumentation acts (e.g. elaborate, clarify). The latter are complex, domain-dependent acts, that can take whole conversations to complete and could be regarded as the dialogue games as we use them in our thesis (see section on global dialogue games below). The following table
lists the four types in increasing order of complexity10(Poesio and Traum, 1997):
Discourse Level Act Type Example
Sub Utterance Unit Turn-taking take-turn, keep turn, release-turn
Utterance Unit Grounding initiate, continue, acknowledge,
repair, request-repair, request-
acknowledge
Discourse Unit Core Speech Act inform, yes-no-question, evaluate,
suggest, request, accept, reject
Multiple DUs Argumentation Acts elaborate, summarize, clarity,
question-and-answer, convince To represent the occurrence of conversation acts in the common ground, Poesio and Traum use a reinterpretation of Muskens’ compositional drt (Muskens, 1994), where DRSs representing the propositional content of con- versation acts are seen as transitions between dialogue situations.
Poesio and Traum (1998) continue their proposal on the treatment of conversation acts in the representation and dynamics of the dialogue situ- ation (now called conversational score), with emphasis on aspects not stud- ied in previous works on speech acts: grounding and participants’ obliga- tions. The paper focuses on core speech acts and grounding acts, using a task-independent dialogue act taxonomy based on the Discourse Resource Initiative (1997) (dri):
• Locutionary acts, representing the utterances of speakers.
• Core speech acts, further classified in terms of their effects on social attitudes (e.g. obligations) to be addressed in the conversation:
– forward-looking function: introduce new social commit- ments (obligations). Obligations can be imposed on the speaker (e.g. commit) or on the hearer (e.g. info-request) and can be conditional (e.g. offer-accept).
– backward-looking function: a response to previous acts
(e.g. accept, answer). Usually discharge obligations, but can
also introduce new ones (e.g. answer introduces an obligation for the hearer to acknowledge the answer).
• Grounding acts: have the effect of moving information from ungroun- ded to grounded sections of the conversational score.
Each dialogue participant keeps a conversational score: a record of the dialogue acts, as well as public beliefs, intentions and social commitments of each participant, with grounded and ungrounded information. Each act leads to an update of the conversational score with one or more discourse units. Discourse units are are first added to the score as ungrounded units. Grounding acts cause a transfer of these units to the grounded section of the score. Forward-looking acts typically introduce discourse obligations once they are grounded. In contrast, appropriate backward-looking acts discharge obligations. These effects are specified by means of update rules with preconditions on the current state of the conversational score and effects representing the changes.
Matheson et al. (2000) implemented part of PTT in the context of the TRINDI project using the TrindiKit dialogue engine (Larsson et al., 2000;
Larsson and Traum, 2000) and information states update rules (Traum and
Larsson, 2003). The theory was implemented as the EDIS dialogue sys-
tem with a focus on dialogue management, more specifically, on aspects of grounding and on the management of obligations. Discourse obligations are tied to dialogue acts by means of update rules. Obligations are introduced by applying these rules when the specified dialogue acts are grounded in in- formation states that satisfy their preconditions. An obligation is discharged when the dialogue act it refers to appears, grounded, on the dialogue his-
tory. The system uses intentions to represent speaker goals, but maximum cooperation is assumed: the system always chooses to meet its obligations before paying attention to its goals. This accounts for behaviours like those that initially motivated Traum and Allen’s (1994) introduction of discourse obligations: e.g. participants responding at all to questions they do not wish to answer. This policy also means that non-cooperative behaviour is precluded, as participants must act against their wishes whenever their obligations are in conflict with their goals.
Obligation-Driven Dialogue Modelling
The modelling approach proposed in this thesis is closest to that taken by Kreutel and Matheson (1999;2000;2003a;2003b). In a series of papers and building on Poesio and Traum’s theory, the authors focus their attention on discourse obligations and show how several of the structures used as primitives in other approaches, such as Grosz and Sidner’s (1986) intentions, Carletta et al.’s (1997) conversational games and Ginzburg’s (1996; 1997) question under discussion, can be derived from a model based on obligations. In addition, the approach is able to account for phenomena left unexplained by these approaches or that require controversial assumptions, such as Boella et al.’s (1999) proposition that cooperation results form the assumption that participants aim at all times to avoid offending each other so that the dialogue progresses smoothly.
Kreutel and Matheson (2001) explore the suitability of obligation-driven modelling for explaining strategic behaviour in discussion scenarios, as op- posed to cooperative behaviour in which the participants assume each other’s intentions to motivate their own actions. The authors illustrate their model with a few handcrafted examples and do not attempt an empirical study as
the one we present in Chapter4.
Kreutel and Matheson regard conversational cooperation essentially in the same way we do in this thesis: participants are conversationally cooper- ative if they act according to their obligations. They refer to cooperation at the level of intentions as traditional cooperation and this is where their model accounts for strategic acting. In other words, participants in their model are conversationally cooperative, contrary to where the focus of this thesis rests. We will come back to this in the Section 2.4.