What are argumentation frameworks in AI?

Argumentation frameworks in AI are formal models used to represent and resolve conflicts in reasoning. They provide a structured way to analyze competing claims by treating arguments as abstract entities and defining relationships between them, such as which arguments attack or support each other. Developed from foundational work by researchers like Phan Minh Dung in the 1990s, these frameworks are often visualized as graphs where nodes represent arguments and edges represent attacks. The goal is to determine which arguments are “acceptable” or justified based on their interactions, enabling systems to make rational decisions even when information is inconsistent.

At their core, argumentation frameworks rely on evaluating sets of arguments that can coexist without internal conflict. For example, consider three arguments: A (“The suspect was at the scene”), B (“The suspect has an alibi”), and C (“The alibi is unreliable”). If A attacks B, and B attacks C, the framework evaluates which subset of arguments forms a logically consistent “extension.” Common evaluation methods include grounded semantics (which prioritizes skepticism) and preferred semantics (which maximizes accepted arguments). In this case, grounded semantics might accept only A and C, while preferred semantics could accept either {A, C} or {B}, depending on the structure. Developers can implement these semantics using algorithms that traverse the graph to identify valid extensions.

These frameworks are applied in AI systems requiring transparent decision-making, such as legal reasoning, chatbots, or autonomous agents. For instance, in a legal AI tool, arguments could represent evidence for or against a defendant, with attacks modeling contradictions. The framework would then output the most defensible conclusions. Similarly, a medical diagnostic system might use argumentation to weigh conflicting symptoms or test results. Developers often use logic programming languages like Prolog or dedicated libraries (e.g., ASPARTIX) to build these systems. While basic frameworks focus on abstract attack relations, extensions like weighted argumentation add numerical strengths to model real-world uncertainty, making them adaptable to complex scenarios.