A Bayesian Agent-Based Framework for Argument Exchange Across Networks

TL;DR

NormAN introduces a normative Bayesian ABM framework for argument exchange across social networks, addressing the gap between content-rich argumentation and population-level belief dynamics. It grounds world structure and evidence in Bayesian networks, enabling agents to update beliefs via Bayes' rule while exchanging arguments under configurable sharing rules. Two case studies illustrate a shift to extremity and polarization versus convergence, showing how evidence distributions and communication norms shape collective outcomes. The work argues that a normative, Bayesian modelling approach provides principled insights into real-world debates and offers an extensible, open-source platform for future research in argumentation and computational social science.

Abstract

In this paper, we introduce a new framework for modelling the exchange of multiple arguments across agents in a social network. To date, most modelling work concerned with opinion dynamics, testimony, or communication across social networks has involved only the simulated exchange of a single opinion or single claim. By contrast, real-world debate involves the provision of numerous individual arguments relevant to such an opinion. This may include arguments both for and against, and arguments varying in strength. This prompts the need for appropriate aggregation rules for combining diverse evidence as well as rules for communication. Here, we draw on the Bayesian framework to create an agent-based modelling environment that allows the study of belief dynamics across complex domains characterised by Bayesian Networks. Initial case studies illustrate the scope of the framework.

Figures (20)

• Figure 1: Illustration of the main components of NormAN. A model within the NormAN framework specifies (1) a ground truth world, (2) a social network, and (3) individual agents communicating across that network. Each of these three components has core specifications. In addition, the square in the middle "activity levels" refers to aspects that could variously be conceived of as properties of world, network or agent. The assignment of model parameters in the current version of NormAN (1.0) to these various model aspects is described in the text.
• Figure 2: The original 'Asia' lung cancer network Lauritzen1988. The Asia BN model was accessed via the bnlearn Bayesian Network Repository (https://www.bnlearn.com/bnrepository/discrete-small.html); it is also one of the exemplar BNs used in the bnlearn package documentation (https://www.bnlearn.com/documentation/bnlearn-manual.pdf). This BN was constructed using a hypothetical case of qualitative medical knowledge to illustrate the utility of Bayes' rule for expert systems. The target hypothesis is ‘Lung’ (whether or not lung cancer is true, shown as the blue node), and there are seven observable evidence nodes (shown as the orange nodes): Asia (recent visit to Asia); smoking; tuberculosis; bronchitis; dyspnoea (shortness of breath); and, x-ray (chest x-ray). The likelihood of lung cancer is increased when smoking, x-ray, bronchitis and dyspnoea are set to true. Combinations of evidence lead to interactions. The ‘either’ node is a deterministic node that is used in this network to represent the fact that both tuberculosis and lung cancer can result in positive x-ray. Network properties: Number of nodes = 8, Number of arcs = 8, Number of parameters = 18, Average Markov blanket size: 2.5, Average degree = 2 and Maximum in-degree = 2.
• Figure 3: Different instantiations of the 'world' defined by the 'Asia' lung cancer network Lauritzen1988. On a given model run, the base net (a), can give rise to different 'worlds' with varying arguments 'for' (green) and 'against' (red), depending on the stochastic initialisation.
• Figure 4: Two groups of 50 agents connected in a social network: a 'small-world' network on the left, and a complete network on the right. Green triangles represent agents who currently support the hypothesis, and red those who do not (cf. Section \ref{['section: Process']}). Both network types are used in the case studies (sections \ref{['shifttoextremity']}, \ref{['polarizationversusconvergence']}). Parameters of the small-world network: rewiring-probability=0.2, k=2.
• Figure 5: Results from simulation using the 'Big Net' network. Shown are the mean beliefs of agents in the target hypothesis and standard error bars, across all 800 experiment runs, at the pre (0) and post-deliberation phase (in this experiment after 25 exchanges/steps). Groups are split by the agents' mean initial direction of belief for a given run.