Abstract：For the same scenario, when without interaction, different agents may have different sets of knowledge about the scenario and reach different conclusions. For example, in the 2020 U.S. presidential election, let’s assume there is a committee consisting of four experts with different backgrounds. They may have their individual observations and understandings towards the natural language arguments and the relations between pairs of arguments, which are presented by two candidates in the presidential debate. Their observations can be represented as different sets of knowledge. Moreover, in the case of non-interaction, four experts are required to reason with the knowledge independently. In terms of formal argumentation, the main research questions of collective argumentation are as follows. How to represent different observations and understandings of different agents towards arguments and the relations between them? How to represent the uncertainties of agents’ beliefs and their individual preferences? How to obtain a collective reasoning outcome to reflect the consensus of the group by aggregating individual reasoning processes and outcomes? How to justify whether the collective outcome is equipped with certain social rationalities with a couple of postulates? With these questions in mind, it is clear that as an emerging research area, the concepts, theories and methodologies of collective argumentation are far from mature.On the basis of existing theories of formal argumentation, this paper divides the existing literature of collective argumentation into two lines of work. We elaborate the main theories and methods on both lines, illustrating their characteristics and drawbacks. In the first line of work, namely the research based on abstract argumentation frameworks, there are two directions: framework merging and semantic aggregation. Framework merging is an operation which first merges individual frameworks with a certain procedure and then computes the semantics of the resulting collective framework(s) to obtain the collective reasoning outcome. Meanwhile, semantic aggregation is an operation which according to a certain rule obtains the collective reasoning outcome directly from individual reasoning outcomes. The difference between these two directions lies in that for the former both knowledge representation and reasoning chain leading to the outcome are available at the collective level, while for the latter only the reasoning outcome is accessible. We adopt Dunne’s axiomatic system to evaluate most of the methods introduced in both lines. The result shows that different methods are equipped with different social rationalities which diverge with each other on the condition that different postulates are satisfied.The second line of work of collective argumentation involves the research on extended argumentation frameworks, such as preference-based argumentation frameworks and probabilistic argumentation frameworks, which include degrees of beliefs and individual preferences into argumentation frameworks. Correspondingly, one may define them as preference-based collective argumentation and probabilistic collective argumentation. Currently, the research in this line is relatively rare and there is no systematic study. For instance, when the properties of a model are studied, degrees of beliefs and individual preferences are usually given in advance and enforced with strong restrictions. Further, no axiomatic system for the evaluation has been investigated.This paper clarifies some important concepts and the scopes of their usages when we investigate and categorize the theories and methods in the existing literature of collective argumentation, which makes the boundary of this research area clearer. Meanwhile, we propose some future work of this area based on the comparison of some existing theories and methods, as well as the analysis on their characteristics and drawbacks. The theories and approaches of collective argumentation can be widely applied to various areas of the field of artificial intelligence, such as smart court, on-line democracy, market prediction, opinion mining, etc.
2022, Vol. 52 
