RESOLVING CONFLICTS BETWEEN PARTICIPANTS IN EXPLANATION GENERATION: direct negotiation vs appeal to authority

Explanation is an informative message that a Speaker S addresses to a Hearer H with several, often intermingled purposes, such as to justify a decision taken, to clarify the meaning of terms or to instruct on how to perform some complex action. The explanation content is thus a function of its communicative goals, of the way that these goals can be reached (the "discourse strategies") and of the context in which the message is given (subjects that have to be illustrated and characteristics of the Hearer). The S's and the H's points of view on these questions are not necessarily the same: H's objectives can be slightly or substantially different from those of S, the explanation strategies that H would use if he or she had the opportunity of participating to generate explanations may be different and, finally, the H's image of the context may be different as well. For example: if the explanation is aimed at persuading H to perform an action to which S attaches a great value, S may be tempted to hide or minimize negative consequences of this action by highlighting positive aspects. On the contrary, H is usually more interested to a more "objective" information. Both agents can under or overestimate the other agent's knowledge; consequently, S might omit relevant items or add redundant information. Finally, some divergence may exist about the way in which specific effects can be obtained: some Ss can have the tendency to adopt an authoritative tone to give instructions to their staff, whereas these Hs would probably prefer a more formal or polite style. In this paper, we consider the case of a text generator that supports the explanation activity of the Speaker. The generation process considers, at the same time, the points of view of S and H, to produce a message being a compromise between them; this compromise can be reached by examining two alternative models of the situation: a. the system behaves as the central authority which, by knowing S's and H's points of view, takes a decision (the system as a Judge); b. the system does its best to solve conflicts between the two agents, through a simulated negotiation process (the system as a Mediator). We studied this problem in the scope of OPADE, a European Community Project aimed at prototyping a system to support the prescription of drugs. We will discuss the limits of the "centralized" approach, to indicate how we are working in the direction of the multiagent alternative.