Modelling Uncertainty in Persuasion

Participants in argumentation often have some doubts in their arguments and/or the arguments of the other participants. In this paper, we model uncertainty in beliefs using a probability distribution over models of the language, and use this to identify which are good arguments (i.e. those with support with a probability on or above a threshold). We then investigate three strategies for participants in dialogical argumentation that use this uncertainty information. The first is an exhaustive strategy for presenting a participant’s good arguments, the second is a refinement of the first that selects the good arguments that are also good arguments for the opponent, and the third selects any argument as long as it is a good argument for the opponent. We show that the advantage of the second strategy is that on average it results in shorter dialogues than the first strategy, and the advantage of the third strategy is that under some general circumstances the participant can always win the dialogue.