Pragmatics First: Coherence in Dialogue Follows Acting in the Environment
Bernd Ludwig · University of Regensburg Publication Server (University of Regensburg) · 2006
Integration of new utterances into context is a central task in any model for rational (human-machine) dialogues in natural language. In this paper, a pragmatics-first approach to specifying the meaning of utterances in terms of plans is presented. These plans are computed during a dialogue on the basis of information about the current situation that is updated continually. A rational dialogue is driven by the reaction of dialogue participants on how they find their expectations on changes in the environment satisfied by their observations of the outcome of performed actions. We present a computational model for this view on dialogues and illustrate it with examples from a real-world application. 1 A Pragmatics-First View on Dialogues Rational dialogues that are based on GRICE’s maxims of conversation serve for jointly executing a task in the domain of discourse (called the application domain) by following a plan that could solve the task assigned to the participants of the dialogue. Therefore, the interpretation of new contributions and their integration into a dialogue is controlled by global factors (e.g. the assumption that all dialogue participants behave in a cooperative manner and work effectively towards the completion of a joint task) as well as by local factors (e.g. how does the new contribution serve in completing the current shared plan?). A robust computational approach that implements such a model of dialogue has to make explicit as many of the above mentioned factors. Only if the factors are represented in an effective and efficient formal language, dialogue systems can be implemented. Examples of such models and their implementation are the slot-filling approach (many state-of-the-art systems implement such attribute-value models), the information-state-update approach (an implemented system is described in [14]), or – more linguistically oriented – approaches like the adjacency-pair models or intentional models such as GROSZ and SIDNER’s (see [10]). Even if it has been noted often that discourse structure and task structure are not isomorphic, only a few contributions to dialogue research focus on the question of how both structures interfere (see Sect. 2). In this paper, we emphasize that it is important to distinguish between the dialogue situation and the application situation: The former is modified whenever speech acts are performed, whereas the latter changes according to the effects of each action being executed. In this section, we will use a MAPTASK dialogue to show what the notions dialogue situation and application situation intend to mean. After presenting related work in Sect. 2, we present our 1 Chair for Artificial Intelligence, University of ErlangenNurnberg, Am Weichselgarten 9, D-91058 Erlangen, email: [email protected] approach first informally in Sect. 3 and then formally by explaining which AI algorithms (Sect. 4 and Sect. 5) we apply in order to turn the informal model into a computationally tractable one. In Sect. 6 we analyze the implications of the proposed method on dialogue and pragmatic strategies for error recovery. 1.1 Talking about Domain Situations The main hypothesis of this paper is that modifications of the dialogue situation are triggered by changes of the application situation. As a response to a speech act, dialogue participants perform a series of actions aiming at achieving some goal. If these actions can be executed, the reaction can signal success. At this point, our understanding of the role of shared plans exceeds that of [9]: GROSZ and KRAUS define an action to be resolved if it is assumed that an agent is able to execute the action. However, in order to understand coherence relations in complex dialogues, it is important to know whether an action has actually been executed and what effect it has produced. Consider the following excerpt from a MAPTASK dialogue (MAP 9, quoted from [6]): R: ++ and ++ you are not quite horizontal you are taking a slight curve up towards um the swamp ++ not obviously going into it G: well sorry I have not got a swamp R: you have not got a swamp? G: no R: OK G: start again from the palm beach G has failed to find the swamp, which means G has failed to perform the action necessary to perform the next one (take a slight curve). In order to solve the current task, R has been able to organize a solution for the task at hand which may or may not involve the other dialogue participant G. How can R put his solution into action? First, he executes each step and, second, validates after each step whether all expectations related to it are fulfilled. 1.2 Talking about Error and Failure In the example above, R’s expectations are not met because G does not find the swamp on the map. However, this would be a precondition for R to continue putting the solution into action that he has organized. On the other hand, G understands that finding the swamp is very important in the current task, but he missed to reach that goal. In order to share this information with R, G verbalizes his failure diagnosis: “I have not got a swamp.” This turn makes R realize that his solution does not work. Obviously, R believed his solution to be well elaborated because he tries to get a confirmation of its failure by asking back “you have not got a swamp?” G’s reacknowledgement is a clear indication for R that it is necessary to reorganize his solution for the current task. Being a collaborative dialogue participant, he will try to recover from that failure to explain the way to the destination. 1.3 Domain and Discourse Strategies For the purpose of recovery, the dialogue participants try to apply a repair strategy that helps them to reorganize the solution. Repair strategies are complex domain dependent processes of modifying tasks and solutions to them. Even being domain dependent in detail, there are some strategies that are domain independent and are regularly adapted to particular domains: • Delay: Maybe it is the best decision to wait a bit and try the failed step again. • Delegation: Maybe someone else can perform better. • Replanning: Maybe another solution can be found taking the current error diagnosis into account. • Relaxation: Modify some parameters or constraints of the task so that a tractable solution can be found. • New Tools: Maybe somehow the dialogue participant can extend his capabilities in the domain so that he can achieve the solution using other, more, or stronger tools and means. • Negotiation: Maybe it is possible to retrieve new helpful information from the user or to come to an agreement of how the task can be modified. • Cancellation: Sometimes giving up to find a solution is the only remaining possibility. This list is necessarily incomplete as depending on the particular domain and current situation in which a dialogue participant has to act these strategies appear in very different fashion. So, it is hard to decide whether exception handling for a single case is taking place or if a particular strategy is being applied. In the example dialogue, G tries to suggest a replanning by telling to R up to what point he was able to understand R’s explanations. According to his communication strategy, a dialogue participant tells his deliberations in more or less detail, sometimes even not at all. This is the case in the example dialogue above. In the last turn, G does not tell that he wants R to reorganize his solution. R must infer this from the content, in particular from the request to restart the explanation at a point that has been passed before the G had failed to understand a step in R’s explanation. This example shows that domain strategies and communication strategies interfere in a dialogue and that complicated reasoning is necessary to identify them in order to react appropriately. Our analysis shows that the notion of coherence is strongly related with the execution of single steps in a solution. Often, coherence cannot be explained satisfactorily within a discourse, but the current situation in which an utterance is made, must be taken into consideration as well.