Multi-Party Argument from Experience

A framework, PISA, for conducting dialogues to resolve disputes concerning the correct categorisation of particular cases, is described. Unlike previous systems to conduct such dialogues, which have typically involved only two agents, PISA allows any number of agents to take part, facilitating discussion of cases which permit many possible categorizations. A particular feature of the framework is that the agents argue directly from individual repositories of experiences rather than from a previously engineered knowledge base, as is the usual case, and so the knowledge engineering bottleneck is avoided. Argument from experience is enabled by real time data-mining conducted by individual agents to find reasons to support their viewpoints, and critique the arguments of other parties. Multiparty dialogues raise a number of significant issues, necessitating appropriate design choices. The paper describes how these issues were resolved in the protocol and implemented in PISA, and illustrates the operation of PISA using an example based of a dataset relating to nursery provision. Finally some experiments comparing PISA with other classifiers are reported. keywords Argumentation, Data mining, Dialogue Games.