💬 NLP

TruthSplit: Operationalizing Conditional Validity in Arguments Through Multi-Perspective Reasoning

The paper presents TruthSplit, an interactive system that operationalizes "conditional validity" by analyzing how arguments yield different conclusions across diverse worldviews through a multi-layer NLI approach and LLM reasoning conditioned on structured value profiles.

Original authors: Benjamin Stieger, Maximilian Terberger, Thomas Huber, Christina Niklaus

Published 2026-06-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Benjamin Stieger, Maximilian Terberger, Thomas Huber, Christina Niklaus

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 watching two people argue about whether a new park should be built in their neighborhood.

Person A says: "We should build it! It will give families a safe place to play and bring the community together."
Person B says: "We shouldn't build it! It will cost too much money and take away space where people could build their own homes."

Both people might be looking at the exact same facts: the cost of the land, the size of the plot, and the number of families in the area. Yet, they reach opposite conclusions. Traditional argument-checking tools usually try to figure out who is "right" or who is making a logical mistake. They ask, "Is the math right? Is the grammar good?"

TruthSplit is a new computer system that asks a different question: "Why do they see the same facts so differently?"

Here is how it works, broken down into simple concepts:

1. The "Worldview Glasses" Analogy

Think of every person wearing a pair of invisible glasses that color how they see the world. These glasses are made of their values (what they care about most), their assumptions (what they believe is true without proof), and their definitions (what words like "freedom" or "justice" mean to them).

Person A might be wearing "Community Glasses." To them, "freedom" means having the support of neighbors.
Person B might be wearing "Individual Liberty Glasses." To them, "freedom" means no one telling them what to do with their own money.

Most computer programs ignore these glasses. They just look at the words. TruthSplit is special because it puts on different pairs of glasses for you. It has pre-programmed "profiles" for six different ways of seeing the world (like a Libertarian, a Religious Conservative, an Eco-Socialist, etc.).

2. How the System Works (The Three-Layer Cake)

When you feed an argument into TruthSplit, it doesn't just read it; it slices it up and analyzes it in three layers:

Layer 1: The Fact Check (The Skeleton)
First, it strips away all the opinions and looks only at the raw logic. Does the evidence actually support the conclusion? If the argument is broken here (like saying "It's raining, therefore I should eat pizza"), the system flags it as a bad argument, no matter who is reading it.
Layer 2: The Perspective Check (The Glasses)
Next, it puts on the "glasses" of a specific worldview. It asks: "If you are a Libertarian, does this argument make sense?" It might find that the argument is perfect for a Socialist but completely wrong for a Libertarian.
Layer 3: The Clash Check (The Map)
Finally, it compares the different worldviews side-by-side. It draws a map showing exactly where they agree and where they fight.
- Example: They might agree that "poverty is bad" (Common Ground), but disagree on whether "taxing the rich" is the solution (Value Conflict).

3. The "TruthSplit" Dashboard

The system gives you a visual dashboard. Imagine a traffic light system for arguments:

Green: This part of the argument is logical.
Yellow: This part makes sense if you believe X, but not if you believe Y.
Red: This is a fundamental clash of values.

It also includes a Chatbot that lets you talk to one of these worldviews. You can ask the "Libertarian" bot, "Why do you oppose this park?" and it will answer strictly using the values and logic of a Libertarian, helping you understand their perspective without them actually being there.

4. What the Researchers Found

The team tested this system with students and experts.

It works: People found it easy to use and could understand why different groups disagree.
It's not a judge: The system doesn't tell you who is "right." Instead, it shows you that the disagreement isn't because one side is stupid or lying; it's because they are starting from different places.
It's consistent: Whether the system used a small computer or a giant AI, the results were the same. The structure of the system mattered more than the specific AI brain used.

The Bottom Line

TruthSplit is like a translator for political arguments. It doesn't solve the argument for you. Instead, it takes the messy, confusing fight and says, "Look, here is exactly where your values clash with theirs. Here is where you define 'freedom' differently. Now that you see the map, you can have a better conversation."

It turns a shouting match into a structured exploration of why we disagree.

Technical Summary: TruthSplit

Problem Statement

Existing computational argumentation tools primarily focus on analyzing intrinsic properties of arguments, such as structure, quality, stance, or persuasiveness. These systems often implicitly assume a universal standard of correctness, classifying arguments as either "valid" or "fallacious." However, in polarized environments, disagreement frequently stems not from flawed reasoning but from deeper structural differences: divergent assumptions about how the world works, distinct value priorities, and varying definitions of contested societal concepts (e.g., "freedom" or "justice").

Current tools fail to address conditional validity: the concept that the validity of a conclusion depends not only on its premises but also on the specific worldview (values, assumptions, and conceptual definitions) through which it is interpreted. Consequently, an argument may be logically consistent and valid within one ideological framework while being invalid in another. There is a lack of systems capable of operationalizing this perspective-dependent analysis to explain why the same claim leads to different conclusions across worldviews.

Methodology

TruthSplit is an interactive system designed to operationalize conditional validity through a multi-stage analysis pipeline. It decomposes arguments and evaluates them against structured, computationally encoded worldview profiles.

1. Worldview Knowledge Base

The system utilizes a structured knowledge base containing six ideological perspectives: Libertarian, Religious-Conservative, Ecological Social-Democrat, Populist-Nationalist, Communist, and Neo-Reactionary.

Structure: Each profile is an extensible JSON object containing weighted core values, key definitions (interpreting contested terms), assumed principles, decision frameworks, and factor scores across 16 ideological dimensions.
Function: These profiles serve as the "priors" that condition the reasoning process, allowing the system to simulate how different perspectives interpret the same input.

2. Analysis Pipeline

The system processes input text (via direct entry, file upload, or news API) through six sequential stages:

Argument Extraction: The system decomposes text into claims (central assertions), premises (explicit evidence), and assumptions (unstated beliefs). This is performed via two modes:
- Local: A sequence classification model (approx. 75–80% accuracy) for privacy-preserving analysis.
- Cloud: An LLM with structured prompts and JSON validation (approx. 95%+ accuracy).
Three-Layer Consistency Testing (NLI): The system employs a Natural Language Inference (NLI) approach to assess logical coherence at three levels:
1. Premise-Claim Logic: Evaluates if premises logically support the conclusion independent of value judgments (perspective-invariant).
2. Perspective-Internal: Assesses if the claim aligns with the principles of a specific worldview profile.
3. Cross-Perspective: Compares consistency scores across profiles to identify fundamental disagreements.
Concept Linking: The system calculates cosine similarity between worldview-specific concept definitions and the extracted claims to identify which concepts are most relevant to each perspective (e.g., how "freedom" is defined differently by Libertarians vs. Social Democrats).
Worldview Reasoning: Using the consistency scores, linked concepts, and full worldview profiles as context, an LLM generates perspective-specific outputs. These include:
- Interpretation and stance (Support/Oppose/Conditional).
- Reasoning chains.
- Key assumptions and concerns.
- Alternative approaches.
Divergence Analysis: The system analyzes how perspectives disagree, categorizing conflicts into:
- Value Conflicts: Competing priorities (e.g., liberty vs. equality).
- Definition Differences: Divergent interpretations of the same term.
- Assumption Gaps: Reliance on different empirical or normative premises.
- Priority Differences: Shared values ranked differently.
- Convergence Flow: A visualization tracing the reasoning chain from core values to conclusions to pinpoint exactly where perspectives diverge.

3. User Interface

The frontend is a React application featuring:

Analysis Dashboard: Visualizes summary metrics, per-worldview reasoning, divergence patterns, and a 3D factor space for worldview positioning.
Worldview Chatbot: An interactive LLM-based agent that embodies a selected worldview, allowing users to engage in dialogue, challenge positions, or explore responses to new scenarios while maintaining strict adherence to the profile's values and definitions.

Key Contributions

The paper outlines four primary contributions:

Operationalizing Conditional Validity: Enabling systematic analysis of arguments across multiple perspectives rather than assigning a single correctness label.
Perspective-Conditioned Reasoning: Generating explicit reasoning chains conditioned on structured, extensible ideological profiles rather than informal prompts.
Interactive Exploration of Divergence: Providing visual and analytical tools to identify value conflicts, assumption gaps, and points of disagreement.
Structured Worldview Representation: Encoding perspectives as computational profiles (values, definitions, decision principles) to support multi-perspective analysis, moving beyond simple text classification.

Evaluation Results

The authors conducted a mixed-methods evaluation involving an expert tier (3 philosophy students) and a broader tier (52 participants from various academic backgrounds).

Usability and Accessibility: Experts rated the system highly for ease of use (4.67/5) and visual appeal (5.00/5). Non-experts found the divergence analysis accessible (4.07/5), indicating the comparative outputs are understandable without specialized philosophical training.
Worldview Representation: Expert validation of 96 factor-worldview combinations showed a moderate positive correlation ( $r=0.33$ ) between implemented factor scores and expert importance ratings. Alignment was strongest for Religious-Conservative and Ecological Social-Democrat profiles. Experts noted inherent ambiguity in quantifying ideological dimensions, with substantial inter-expert variance.
LLM Robustness: Outputs generated by different LLM families (Claude, GPT, Gemini, Grok, DeepSeek) showed no significant quality differences, suggesting the structured prompting approach standardizes reasoning across models.
Claim Extraction: Rated moderately positively (6.67/10), indicating room for improvement in automated extraction accuracy.

Significance and Claims

The paper positions TruthSplit not as a tool to resolve disagreement, but as an analytical and educational instrument that makes ideological disagreement more intelligible.

Reframing Disagreement: By foregrounding conditional validity, the system reframes conflict as a product of differing premises and value hierarchies rather than factual inconsistency or logical failure.
Beyond Binary Judgments: It demonstrates how AI systems can move beyond binary "correct/incorrect" judgments toward a transparent analysis of normative disagreement.
Educational Utility: The system is intended for educators, students, researchers, and analysts to explore perspective-dependent reasoning in political and ethical discourse.
Modest Scope: The authors acknowledge limitations, including the exploratory nature of the study, the limited sample size, and the challenge of defining rigid boundaries for worldviews. They explicitly state that the evaluation focuses on the usability and interpretability of the analyses, not the absolute correctness of the reasoning or divergence explanations, leaving the latter for future work.

The authors conclude that structured worldview conditioning provides a viable foundation for multi-perspective argument analysis, offering a framework for more informed engagement with ideological differences.