🤖 AI

Rejecting Arguments Based on Doubt in Structured Bipolar Argumentation

This paper introduces Structured Bipolar Argumentation Frameworks (SBAFs) and their associated semantics to model rational agents who can reject arguments based on mere doubt and evaluate acceptability at the sentence level, thereby bridging the gap between admissible and complete semantics while offering a generalized perspective on existing approaches like deductive support.

Original authors: Michael A. Müller, Srdjan Vesic, Bruno Yun

Published 2026-02-04

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Michael A. Müller, Srdjan Vesic, Bruno Yun

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are sitting in a town hall meeting where people are shouting out arguments to convince you of something. Usually, computer programs designed to analyze these debates work like a strict, rigid judge. They follow a simple rule: "If an argument is defended and nobody has successfully shot it down, you must accept it."

This paper proposes a different kind of judge—one that acts more like a skeptical human being. It introduces a new way for computers to handle debates that allows for doubt and focuses on individual facts rather than just whole arguments.

Here is a breakdown of the paper's ideas using simple analogies:

1. The Problem: The "Rigid Judge" vs. The "Skeptical Human"

In traditional computer argumentation, if you have a chain of logic that no one has attacked, the computer says, "This is true, accept it."

The Paper's View: Humans aren't like that. Even if an argument hasn't been attacked, you might still reject it if you doubt its starting point.
The Analogy: Imagine someone says, "This violin is a Stradivarius because Alex said so."
- The Rigid Judge: "Nobody attacked Alex's statement. Therefore, the violin is a Stradivarius. Accept it."
- The Skeptical Human (This Paper): "I don't know Alex. He might be lying or clueless. Even though no one attacked his claim, I doubt it. So, I will reject the conclusion that it's a Stradivarius."

The paper argues that computers should be allowed to say, "I'm not convinced yet," without needing a specific counter-argument to prove them right.

2. The New Tool: "Structured Bipolar Argumentation"

To make this happen, the authors built a new framework called Structured Bipolar Argumentation Frameworks (SBAFs). Think of this as a new rulebook for debates with two special features:

Bipolar (Two Sides): In old systems, arguments only had "attacks" (like punches). This new system adds "supports" (like hand-holding). If Argument A supports Argument B, accepting A makes it harder to reject B.
Structured (The Ingredients): Old systems treated arguments like black boxes. You either accepted the whole box or rejected it. This new system looks inside the box. It separates the premises (the ingredients) from the conclusion (the cake).
- The Analogy: Instead of just saying "I accept the cake," the system asks, "Do you accept the flour? Do you accept the sugar?" You might accept the flour and sugar but still doubt the baker's ability to mix them, so you reject the cake even if you like the ingredients.

3. The Two New Rules: "Coherence" and "Adequacy"

The paper defines two ways to decide what to believe, based on how strict you want to be about doubt.

A. Coherent Argument Extensions (The "Argument" View)

This looks at which arguments you accept.

Weak Coherence: You can reject an argument even if it's defended, as long as you doubt its premises. It's like saying, "I see you have a shield (defense), but I don't trust the sword you're holding (the premise), so I'm not buying it."
Strong Coherence: You are stricter. You only reject an argument if you have specific evidence that the premise is wrong (an "undercut"). If you just suspect it's wrong but have no proof, you have to accept it.

B. Adequate Language Extensions (The "Sentence" View)

This looks at which sentences (facts) you accept, ignoring the arguments for a moment.

The Analogy: Imagine you are filling out a checklist of facts.
- Weak Adequacy: You check off facts you are sure of. If a fact depends on a shaky argument, you leave it unchecked.
- Strong Adequacy: You check off facts unless you have proof they are false.
Why this matters: Sometimes, you might accept a fact (e.g., "Clara said Anne-Sophie owns the violin") but reject the conclusion (e.g., "Anne-Sophie owns the violin") because you doubt the source. This system allows you to say, "I accept the claim that Clara said it, but I don't accept that it's true."

4. The "Sweet Spot"

The authors found that their new rules sit in a "Goldilocks zone" between two existing computer rules:

Admissibility: Too loose. It lets you reject anything you want, even if it's well-supported.
Completeness: Too strict. It forces you to accept everything that isn't attacked.
Their New Semantics: Just right. It forces you to accept supported arguments unless you have a reason to doubt them, but it doesn't force you to accept things you have no evidence for.

5. The Big Discovery: When Does Structure Matter?

The paper runs a fascinating test: When can we ignore the complex structure of arguments and just look at simple attacks?

The Finding: If the debate is "saturated" (meaning every single fact has its own tiny, self-contained argument attached to it), then the complex "sentence-based" view and the simple "argument-based" view give the same results.
The Analogy: If every single brick in a wall has its own name tag, you can just look at the bricks. But if the bricks are glued together in complex patterns, you have to look at the whole wall to understand what's happening.
Implication: In simple cases, the old, simpler computer models work fine. But in complex, real-world debates where premises are shared and mixed, you need this new, more detailed model to get the right answer.

Summary

This paper teaches computers to be more like human debaters. It allows them to:

Doubt an argument without needing a counter-attack.
Separate the facts (sentences) from the logic (arguments).
Choose between being a strict skeptic (Strong Coherence) or a flexible one (Weak Coherence).

It doesn't claim to solve all human disagreements, but it provides a better mathematical tool for modeling how we actually think when we are unsure.

Technical Summary: Rejecting Arguments Based on Doubt in Structured Bipolar Argumentation

Problem Statement

Current computational argumentation frameworks, particularly those based on abstract argumentation (e.g., Dung's semantics), often rely on two assumptions that diverge from informal reasoning practices:

Automatic Acceptance of Defended Arguments: Formal semantics (specifically complete semantics) typically require rational agents to accept all arguments they can defend. In contrast, informal approaches allow agents to reject defended arguments if they harbor doubts about specific premises, even in the absence of a formal counter-argument.
Holistic Argument Evaluation: Formal approaches often treat arguments as atomic units, evaluating them as a whole. Informal approaches suggest that agents naturally evaluate specific sentences or claims within a debate rather than entire argument structures.

Existing bipolar argumentation frameworks (which include support relations) often still rely on completeness-based semantics or deductive support semantics that force the acceptance of supported arguments, failing to capture the nuance of "doubt" where an agent accepts premises but rejects the inference or the conclusion.

Methodology

The authors propose Structured Bipolar Argumentation Frameworks (SBAFs) to integrate structured argumentation (premises and conclusions) with bipolar relations (attacks and supports).

1. Framework Definition

Language ( $L$ ): Defined as a tuple $\langle L, \bowtie, n \rangle$ , where $L$ is a set of sentences, $\bowtie$ is a symmetric incompatibility function (contrariness), and $n$ is a naming function that maps argument structures to sentences representing inference claims.
Arguments: Defined as tuples $\langle Prem(a), Conc(a) \rangle$ , where $Prem(a)$ is a non-empty finite set of sentences and $Conc(a)$ is a sentence.
Relations:
- Attack ( $\rightarrow$ ): An argument $a$ attacks $b$ if $Conc(a)$ is incompatible with a sentence in $Sent(b)$ (rebuttal) or if $Conc(a)$ is incompatible with the name of $b$ (undercut).
- Support ( $\rightarrow\!\!\!\rightarrow$ ): A set of arguments $E$ supports an argument $a$ if $Prem(a) \subseteq Sent(E)$ . This captures "premise support" and "premise-sharing support" (exhaustion).

2. Semantics

The paper introduces two layers of semantics: Argument Semantics (evaluating sets of arguments) and Language Semantics (evaluating sets of sentences). Both layers offer "Strong" and "Weak" variants to model different interpretations of doubt.

Argument Semantics

The authors define Coherent Argument Extensions as admissible extensions that satisfy specific closure conditions regarding support:

Strong Coherence: An extension $E$ is strongly coherent if it satisfies Strong Support-Closure: If $E$ supports $a$ and $E$ does not contain undercutting information for $a$ , then $a \in E$ . This forces acceptance of supported arguments unless explicitly undercut.
Weak Coherence: An extension $E$ is weakly coherent if it satisfies Weak Support-Closure: If $E$ supports $a$ , $E$ does not contain undercutting information for $a$ , and $E$ defends $a$ , then $a \in E$ . This allows agents to reject supported arguments if they are not defended, effectively modeling doubt about the inference even if premises are accepted.

Language Semantics

To address the evaluation of sentences directly, the authors define Adequate Language Extensions:

Strong Argument Set ($Args(S)$): The set of arguments whose premises are in $S$ and whose names are not in $S$ .
Weak Argument Set ( $Arg_w(S)$ ): Defined via a fixpoint construction analogous to complete extensions, ensuring defended arguments are included.
Strongly/Weakly Adequate Extensions: A set of sentences $S$ is adequate if it is compatible (no incompatible sentences), satisfies sentence-closure (includes conclusions of accepted arguments), and the corresponding argument set (strong or weak) defends all its arguments.

Key Contributions

Formalization of Doubt: The paper formalizes the concept of "critical reaction" where an agent can reject a defended argument based on doubt about a premise or inference, provided there is no support to remove that doubt. This is achieved by relaxing the requirement to accept all defended arguments found in complete semantics.
Dual Perspective (Argument vs. Language): The framework provides semantics for both argument extensions and language extensions. This allows agents to communicate based on accepted sentences rather than complex argument structures, bridging the gap between abstract and structured argumentation.
New Semantics Spectrum: The proposed semantics (Strong/Weak Coherence and Adequacy) occupy a theoretical space between admissible and complete semantics. They are weaker than complete semantics (allowing rejection of defended arguments) but stronger than admissible semantics (respecting support relations).
Reinterpretation of Existing Semantics:
- Preferred Semantics: The paper shows that in saturated frameworks, preferred extensions correspond to "confident weakly coherent" extensions. This implies that maximizing accepted arguments is equivalent to maximizing accepted sentences only under specific structural conditions (strong saturation).
- Deductive Support Semantics: The paper demonstrates that strong coherence coincides with d-admissibility (deductive support semantics) when arguments have exactly one unique premise and no undercutting information exists. This generalizes deductive support to more complex argument structures.

Results and Theoretical Findings

Directionality: Weak coherence satisfies the property of directionality (arguments only affect acceptability if connected via directed paths), whereas strong coherence does not.
Correspondence: In saturated SBAFs, there is a direct correspondence between coherent argument extensions and adequate language extensions.
Maximization Strategies: The strategy of maximizing accepted arguments (preferred extensions) and maximizing accepted sentences (confident weakly coherent extensions) do not always coincide. They coincide only in strongly saturated frameworks where minimal arguments for all sentences are present.
Undercutting: The framework explicitly handles undercutting via the naming function, allowing agents to reject inferences even if premises are accepted, provided the inference claim is attacked.

Significance and Claims

The paper claims to offer a new perspective on computational argumentation informed by informal theories (philosophy, linguistics). Its primary significance lies in:

Bridging Formal and Informal: It incorporates the informal intuition that agents can doubt and reject arguments without a full counter-argument, and that agents reason about sentences directly.
Generalization: It generalizes existing bipolar semantics (specifically deductive support semantics) to structured arguments with complex premises.
Flexibility: By distinguishing between strong and weak interpretations, the framework allows for modeling different levels of skepticism or commitment in agents.
Clarification of Support: It clarifies when abstract argumentation (ignoring support) is warranted (in strongly saturated frameworks with preferred semantics) and when structured support is necessary.

The authors note that their approach implements a "generous notion of doubt," which might make it difficult to force agents to accept arguments. They suggest future work could distinguish between doubtable sentences and a "common ground" of accepted sentences, but do not implement this within the current framework. The paper concludes that SBAFs provide a robust mechanism for evaluating debates where the structure of arguments and the specific sentences involved are critical to the agent's reasoning process.