Relevance in Structured Argumentation

Annemarie Borg, Christian StraßerIJCAI 2018 · DOI

TL;DR

This paper investigates how to ensure that conclusions drawn from structured argumentation systems remain robust even when irrelevant information is added. The authors introduce a general framework for structured argumentation and define key properties like "Non-Interference" and "Cumulativity" to formally capture notions of relevance, showing how these properties can be inherited by various argumentation systems.

What Problem Does This Paper Address?

In artificial intelligence, particularly in areas like knowledge representation and reasoning, we often build systems that draw conclusions from a set of premises. Structured argumentation is a popular approach where arguments are built with logical structure, and their validity is determined by how they interact (e.g., attack each other). A crucial aspect of any reasoning system is its "relevance" – meaning it should ideally ignore irrelevant information. Imagine a legal system: if you present a case, adding unrelated facts shouldn't change the outcome. Similarly, in argumentation systems, we want the conclusions to be stable against the introduction of extraneous information.

The challenge is that many existing structured argumentation frameworks, while powerful, can be sensitive to irrelevant inputs. This sensitivity can lead to unexpected or incorrect conclusions. The authors aim to identify conditions under which structured argumentation systems exhibit desirable "relevance" properties, ensuring their conclusions are robust. They focus on two main types of relevance: syntactic (information sharing propositional variables) and semantic (information that shouldn't have defeating power).

Methods

The paper proposes a general, abstract framework for structured argumentation. This framework is designed to be flexible enough to represent many existing systems (like ASPIC, ABA, logic-based argumentation) while simplifying the meta-theory. Key components of their framework include:

  • A formal language L: The basic building blocks of statements.
  • A deducability relation ├: A relation that defines what can be derived from a set of formulas. Crucially, this relation doesn't assume standard logical properties like monotonicity.
  • Arguments: Defined as premise-conclusion pairs (Γ, γ), where Γ is a finite set of formulas and γ is a formula.
  • Attacks: A specific type of attack is defined where an argument attacks another if its conclusion conflicts with a premise of the other argument. This is a simplification that allows for broader representation.

Using this general setting, they define and investigate several relevance-related properties:

  • Pre-Relevance: A property of the underlying deducability relation that ensures conclusions are robust under adding irrelevant premises.
  • Non-Interference: A syntactic relevance property where adding information that shares no propositional variables with the original information does not change the conclusions.
  • Crash-Resistance: A property ensuring that no set of formulas is "contaminating" (i.e., can arbitrarily change conclusions).
  • Cumulativity: A semantic relevance property that intuitively means adding derivable formulas to the premises doesn't change the set of conclusions.

They then show how these properties can be inherited by different argumentation systems based on their underlying logic and attack definitions.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • A Unified Framework for Relevance: The paper provides a general framework that allows for a unified study of relevance properties across diverse structured argumentation systems. This is valuable because comparing results across different frameworks can be difficult.
  • Syntactic vs. Semantic Relevance: The authors clearly distinguish between syntactic relevance (based on shared propositional variables) and semantic relevance (based on the logical behavior of conclusions). They show how both can be formally captured and studied.
  • Pre-Relevance as a Key Enabler: The concept of "Pre-Relevance" for the underlying deducability relation is identified as a crucial condition for ensuring that structured argumentation systems inherit desirable relevance properties like Non-Interference.
  • Cumulativity for Grounded Semantics: They demonstrate that "pointed" argumentation settings (a specific structural property) are sufficient to guarantee Cumulativity for grounded semantics, a significant result for understanding the stability of conclusions in this common semantics.
  • Representing Existing Systems: The paper shows how popular systems like Logic-Based Argumentation, Assumption-Based Argumentation (ABA), and ASPIC can be translated into their general framework, allowing their theoretical results to be applied to these practical systems.

Key Results

  • Theorem 1: If the underlying deducability relation satisfies Pre-Relevance, then the structured argumentation system satisfies Non-Interference for grounded and preferred semantics.
  • Theorem 2: If an argumentation setting is "pointed" (satisfies Cut and has a specific structure), it satisfies Cumulativity and Extensional Cumulativity for grounded semantics.
  • Theorem 3: If the underlying relation satisfies Cut and is contrapositable, and the setting uses a specific definition of consistency, the system is cumulative for grounded, preferred, and stable semantics.
  • The paper provides detailed mappings of various attack types in logic-based argumentation to their general framework, showing how to simulate them and inherit relevance properties.

Limitations and Future Work

  • Focus on Skeptical Semantics: The paper primarily focuses on skeptical consequence relations (Grd, Prf, Stb) for space reasons, noting that other semantics might behave differently.
  • Non-Prioritized Settings: The current work is restricted to non-prioritized settings. Future work explicitly mentions incorporating priorities, which is a significant extension in argumentation theory.
  • Broader Relevance Properties: The authors suggest extending their study to other meta-theoretic properties beyond Non-Interference and Cumulativity, such as other rationality postulates.
  • Handling Defeasible Premises: While the framework can represent systems with defeasible premises, the examples and proofs sometimes simplify this aspect.

The authors plan to extend their work to include priorities and explore a wider range of non-monotonic inference properties, aiming to provide a more comprehensive understanding of relevance in structured argumentation.

Summary updated

BibTeX
@inproceedings{borg2018relevance,
  author = {Annemarie Borg and Christian Straßer},
  title = {{Relevance in Structured Argumentation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/242},
}
Figure