Use of argumentation theories in practice by the Dutch police

This generator generates ASPIC+ argumentation theories that are similar in structure to the argumentation theory that models the domain knowledge concerning fraud intake at the Dutch police. This use case is described in [2]. Note that, in this use case, the argumentation theory is not used for the "static" problem of acceptance (or listing extensions), but for the "dynamic" problems of stability and relevance.

Properties of generated ASPIC+ argumentation theories

The generator uses generators from PyArg to generate ASPIC+ argumentation theories. Properties of the ASPIC+ argumentation theories:

• Layered structure: rules go from knowledge to "topic" literals.
• The highest layer is 4, so the path from argument premise to conclusion can contain at most 4 defeasible rules.
• Contrariness by classical negation: each literal in the language has exactly one contradictory --- its negation.
• No strict rules, just defeasible rules.
• No ordinary premises, just axioms.
• No preference ordering on the defeasible rules.
• No undercutters.
• All rules have 1, 2, 3, 4 or 5 antecedents.

Conversion into abstract argumentation frameworks

Any ASPIC+ argumentation theory can be converted into an abstract argumentation framework. We do this by listing all arguments inferred by the argumentation theory and identifying all (rebuttal) attacks between those arguments. Some properties of the resulting AFs:

• Long chains of attacks are impossible.
• For any attack (A,B) either A is an axiom or A is attacked by B itself or another argument (which is a subargument of B).

Conversion into ABA frameworks

Alternatively, the ASPIC+ argumentation framework can be converted into an ABA framework. For this, we use the method from Section 7 of [1]. Some properties of the resulting ABA frameworks:

• If the ASPIC+ framework had l literals, then the ABA framework has 4 * l atoms (because the ASPIC+ framework with l literals is created with 1.5 * l rules and ABA introduces two new atoms for every ASPIC+ rule).
• If the ASPIC+ framework had r rules, then the ABA framework has 2 * r rules.
• The ABA framework is flat: no assumption is the conclusion of a rule.

Example usage

To install dependencies, run

pip install -r requirements.txt

This installs PyArg.

The benchmarks can then be generated using the following command.

python .\console.py 50 3 ".\generated" "aba"

The first argument (50) refers to the number of literals for the ASPIC+ argumentation theory or theories. The second argument (3) represents the number of instances that should be created. The third argument (".\generated") refers to the folder where the generated instances should be placed. The fourth argument ("aba") specifies the output format, which can be "aba" or "af".

References

1. Heyninck, Jesse and Christian Strasser. "Relations between assumption-based approaches in nonmonotonic logic and formal argumentation." 16th International Workshop on Non-Monotonic Reasoning. 2016.
2. Odekerken, Daphne, et al. "Approximating stability for applied argument-based inquiry." Intelligent Systems with Applications 16 (2022): 200110.