ITA Framework: An Interpretable Fact-Checking System Integrating Neuro-Symbolic Reasoning

This article introduces the ITA (Inference-Time Argumentation) framework. Addressing the limitations of binary classification (true/false) in fact-checking for high-risk scenarios and the issue of unfaithful post-hoc explanations, it integrates formal argumentation semantics with LLMs to achieve ternary fact-checking (true/false/uncertain). It ensures consistency between prediction results and reasoning processes, balances performance and interpretability, and provides reliable AI support for fields like healthcare and finance.

Real-World Dilemmas

In high-risk scenarios such as healthcare, finance, and law, the accuracy of fact-checking is crucial. However, the information environment is full of uncertainties: incomplete evidence, conflicting sources. Binary classification (true/false) is overly simplistic, and an "uncertain" judgment is more honest; users need to understand the basis for the model's decisions.

Issues with Post-Hoc Explanations

Mainstream LLMs first generate answers and then construct explanations (post-hoc explanations), which easily leads to a disconnect between the explanation and the actual reasoning ("unfaithful explanations"). In high-risk scenarios, users cannot verify the basis of the model's judgments, leading to potential risks of bias or incorrect associations.

Definition of the ITA Framework

Inference-Time Argumentation (ITA) is a ternary fact-checking framework integrating neuro-symbolic reasoning. Its core is to integrate formal argumentation semantics into LLM training and inference.

Role of Formal Argumentation Semantics

1. Quantify argument strength: Assign a base score (intrinsic credibility) to each argument
2. Calculate claim strength: Derive a comprehensive strength based on attack/support relationships between arguments
3. Derive ternary judgment: Map strength to true/false/uncertain

Integration of Neural and Symbolic Components

• Training Phase: LLMs learn to generate arguments and assign scores, with the goal based on the contribution of argument quality to predictions
• Inference Phase: Output is determined via formal semantic calculations, ensuring an inevitable link between predictions and argument structures

Workflow

1. Argument Generation: LLMs analyze claims and evidence to generate a set of supporting/opposing arguments
2. Score Assignment: Assign base scores to arguments that reflect evidence strength and source credibility
3. Interaction Analysis: Identify attack/support relationships between arguments
4. Strength Calculation: Apply formal semantics to compute the comprehensive strength of the claim
5. Ternary Classification: Categorize into true/false/uncertain based on thresholds

The workflow ensures every step is traceable, and users can check the arguments and strength calculation process.

Experimental Results

Evaluated on two ternary fact-checking datasets:

• Comparison with argumentation baselines: Significantly outperforms traditional methods, due to end-to-end training optimizing argument generation and score assignment, more rigorous formal semantics, and neural components handling natural language complexity
• Comparison with direct prediction baselines: Performance is comparable, breaking the assumption that "interpretability and performance cannot coexist"

Interpretability Advantages

• Explicit argument structure: Lists all arguments supporting/opposing the claim
• Computable scores: Argument strength has clear mathematical definitions and calculation processes
• Deterministic reasoning: Given arguments and scores, the conclusion is uniquely determined
• Auditable decisions: Users/auditors can verify the correctness of reasoning step by step

Healthcare Information Fact-Checking

• Mark claims with insufficient evidence as "uncertain"
• Provide argument chains supporting/opposing medical claims
• Help users understand the credibility of health advice

Financial Public Opinion Analysis

• Analyze statements related to a company's financial status
• Distinguish between judgments with sufficient evidence and speculative conclusions
• Provide structured argument references for analysts

Legal Auxiliary Research

• Evaluate arguments about the relevance of precedents to current cases
• Identify evidence gaps in legal claims
• Support lawyers in constructing strong argument structures

Technical Significance

ITA is an important advancement in the field of neuro-symbolic AI, organically combining the expressive power of deep learning with the interpretability of symbolic reasoning, rather than simply opposing them.

Core Points

• Integrate formal argumentation semantics to achieve deep integration of neural and symbolic reasoning
• Ternary output adapts to information uncertainty
• Eliminate the problem of unfaithful post-hoc explanations
• Balance prediction performance and structured, auditable interpretability

Future Directions

• Multilingual expansion: Apply to non-English scenarios
• Real-time argument updates: Support corrections in dynamic evidence environments
• Human-machine collaborative argumentation: Combine expert knowledge to improve automatic arguments
• Adversarial robustness: Enhance stability when facing misleading evidence