A Heuristic Strategy for Persuasion Dialogues

Argument-based dialogues allow agents to effectively communicate both their beliefs and the reasons they have for holding those beliefs; consequently, they have become a useful mechanism for agent co-ordination, particularly in the domains of human-machine interaction and agreement technologies (Modgil et al. 2013). We focus on a simple type of persuasion dialogue (where one agent presents arguments to another with the aim of convincing it to accept some argument that is the topic of the dialogue) and consider the problem of how the persuader can determine which argument to present at each step of the dialogue, i.e., what dialogue strategy it should employ. The development of methods for generating agent dialogue strategies is an active area of research in argumentation (Thimm 2014). So far, work that has investigated this problem has shown that computing an optimal strategy for two-party dialogues is computationally expensive, and becomes intractable as the number of arguments in the dialogue domain increases. Black et al.’s approach (Black, Coles, and Bernardini 2014) considers the same simple persuasion dialogue setting that we focus on here, modelling it as a planning problem so that a planner can be used to generate an optimal strategy for the persuader, while the approaches of Hadoux et al. (Hadoux et al. 2015) and of Rienstra et al. (Rienstra, Thimm, and Oren 2013) each support richer models of argument dialogue, generating optimal strategies using Mixed Observability Markov Decision Problems (MOMDPs) and a variant of the minimax algorithm respectively. While each of these approaches (Black, Coles, and Bernardini 2014; Hadoux et al. 2015; Rienstra, Thimm, and Oren 2013) determines an optimal strategy for the persuader, none have been shown to scale to domains with more than 10 arguments. We present a heuristic strategy for persuasion that we show can easily scale to domains with 50 arguments. A realworld example of such a domain is Decide Madrid1, an online forum in which citizens can participate in debates in order to make meaningful decisions about local government policy; debates on this site commonly have in excess of 50 arguments. Although the heuristic strategy we present is not optimal, we show that it gives a reasonable chance of suc-