Explainable Automated Fact-Checking

• Explainable automated fact-checking is a method that integrates NLP, graph reasoning, and causal inference to verify claims and produce human-interpretable justifications.
• It employs multi-stage rule-based and multimodal pipelines to deliver verdicts alongside detailed explanation metrics like faithfulness, coherence, and actionability.
• The approach advances transparency in misinformation mitigation by balancing prediction accuracy with rigorous, audit-friendly explanations in high-stakes domains.

Explainable automated fact-checking (EAFC) synthesizes natural language processing, graph reasoning, and causal inference to deliver not only a veracity verdict for factual claims but also transparent, interpretable justifications for those decisions. EAFC is distinguished by its dual emphasis on both prediction accuracy and the ability to generate justifications that meet rigorous technical desiderata such as faithfulness, coherence, actionability, and causal traceability. The domain draws upon multi-stage rule-based pipelines, neural explanation generators, graph-based reasoning, question-answering frameworks, and multimodal inference, representing a central paradigm shift in the automation and auditability of truth claims across high-stakes domains.

1. Formal Foundations and Explanation Taxonomy

EAFC builds upon the computational paradigm of automated fact-checking (AFC), in which the system verifies the truthfulness of a textual claim $c$ based on retrieved evidence $E$ and outputs a discrete veracity label $\hat y$ (e.g., Support/Refute/Neutral) with an accompanying human-interpretable justification $\text{Justif}$ (Kotonya et al., 2020, Eldifrawi et al., 2024). Explanations appear in several forms:

• Textual summaries: concise natural language rationales, often via extractive or abstractive summarization (e.g. [Kotonya & Toni 2020, (Kotonya et al., 2020)]).
• Highlighted evidence spans: token- or sentence-level annotations, possibly weighted by model attention or post-hoc attribution (e.g. [Popat et al. 2018, Shu et al. 2019]).
• Graph/rule-based traces: logical inference chains or knowledge graph paths (e.g. [Ahmadi et al., Gad-Elrab et al.]).
• Question-answer tables: aspect-wise comparisons produced by decomposing claims into sub-questions (e.g. (Rani et al., 2023, Yang et al., 2021)).
• Multimodal justifications: combinations of features (metadata, graph features, highlighted words) for social or visual input (Lourenço et al., 11 Aug 2025, Akhtar et al., 2023).

EAFC systems are classified according to the interplay of prediction and explanation modules: joint models that optimize for both tasks, sequential pipelines where explanations follow verdicts, and multi-hop reasoning chains that route explanations through intermediate inferences (Eldifrawi et al., 2024).

2. Rule-Based, Causal, and Multi-Hop Reasoning

Recent advances underscore the importance of explicit causal reasoning, particularly the detection of erroneous cause-effect relationships within claims and corresponding evidence. A representative pipeline first extracts event-relation triples using event extraction (e.g., REBEL) for both claim and evidence, then constructs directed causal graphs $G = (V, L)$ using a fixed ontology of semantic relations $$\mathcal{R} = \{\text{direct-cause}, \text{prevent}, \text{intends-to-cause}, \text{enables}\}$$ (Rebboud et al., 15 Dec 2025). Predicates for semantic similarity and polarity, operationalized via cosine similarity and fine-tuned sentiment analysis, mediate the comparison of events.

The core reasoning engine applies deterministic rule schemas:

• Logical Alignment: If claim and evidence event chains are semantically and causally aligned via direct relations or transitive chains, label as Supported.
• Logical Misalignment: If evidence provides causal information that contradicts the claim (e.g., an event both causes and prevents a target), label as Refuted.
• Causal Loop: If the combined claim/evidence chains form a directed cycle with consistent transitivity properties, label Supported.
• Cherry-Picking Conflict: If evidence contains mutually dissimilar or oppositional event outcomes under the same relation, label as Conflicting.

Every verdict is grounded in a symbolic proof trace, enabling explicit, stepwise demonstration of which event chains, semantic similarities, or polarity mismatches dictate the final decision (Rebboud et al., 15 Dec 2025).

3. Data Modalities and Multimodal Explanation Pipelines

EAFC integrates diverse content types and metadata, extending explainability beyond pure text analysis. Multimodal systems process:

• Textual Features: Embedded by language-specific transformer models (BERTweet, RoBERTuito), with concatenation of shallow metadata (retweet counts, timestamps).
• Graph Features: Social interactivity via graph attention networks (GAT), computing node representations with weighted edge aggregation. Explanations from these features leverage graph-based LIME or HSIC-Lasso for modal-agnostic interpretability, surfacing influential nodes, edge types, or metadata features.
• Text-based Attributions: IG, LIME, or SHAP to attribute prediction decisions to specific input words or tokens.
• Combined Views: Present users with lists of decisive graph connections and highlighted textual evidence (Lourenço et al., 11 Aug 2025).

Experiments on social datasets demonstrate that cross-modal fusion (text + metadata + graph) yields superior accuracy (F1 up to 0.97), and robustness protocols indicate text-based explanations are less susceptible to noisy features (Lourenço et al., 11 Aug 2025).

4. Evaluation Frameworks and Explanation Quality Metrics

EAFC explanation quality is measured via both automatic overlap metrics and human-centered protocols:

• Automatic metrics: ROUGE-1/2/L, BLEU, METEOR, BERTScore, BLEURT for NL explanations (Feher et al., 2024, Akhtar et al., 2023). Precision, recall, F1 for token-level rationale overlap versus annotated gold (Kotonya et al., 2020).
• Faithfulness and coherence: Compute entailment relations between claim, evidence, and generated explanations using pretrained NLI models. Properties include strong global coherence ($\forall_i: e_i \models c$), weak global coherence (no contradiction), and local coherence (no internal contradiction) (Kotonya et al., 2020).
• Human protocols: Assess interpretability, trustworthiness, robustness, convincingness, hallucination, contradiction, coverage, redundancy, and overall quality (Feher et al., 2024, Lourenço et al., 11 Aug 2025).
• Actionability: The FinGrAct framework introduces fine-grained criteria—error detection, correction, supporting reference existence/relevance/support—and aggregates into a Likert-style score strongly correlated with human ratings (Eldifrawi et al., 7 Apr 2025).

High-performing generative models (T5-large, LED-base) reach ROUGE-1 ≈47 on full-text explanations, and metric learning models (DeBERTa-v2-xxlarge) achieve Matthews Correlation Coefficient ≈0.7 on contradiction/hallucination detection (Feher et al., 2024). FinGrAct advances actionability scoring, while CLUE explicitly traces sources of uncertainty by identifying and verbalizing evidence conflicts via attention mining and span-clustering (Sun et al., 23 May 2025).

5. Practitioner Guidance, Human-Centric Criteria, and Limits

Empirical studies and interviews with professional fact-checkers reveal persistent gaps between EAFC system outputs and practitioner needs. Users demand explanations that:

• Show the reasoning path (traceability): Document every step from evidence retrieval to verdict, emulating human fact-check workflows.
• Reference specific, verifiable evidence: Pinpoint sentences, datasets, or images that underpin the model’s decision.
• Highlight uncertainty and gaps: Quantify model confidence, surface missing/conflicting data, and suggest next steps.
• Support auditability and user correction: Link to full sources, enable user feedback to override or refine rationales, and facilitate continual improvements (Warren et al., 13 Feb 2025, Zhang et al., 2021).
• Ensure faithfulness and context-fullness: Explanations must mirror the model’s actual computation, referencing all relevant context—claim, evidence, temporal markers—and avoid hallucinations.

Design patterns for EAFC interfaces include multi-stage dashboards, trace panels, interactive evidence marking, and model data cards documenting limitations and training provenance. Integration of process-centered, user-in-the-loop, and multi-dimensional transparency principles is essential for real-world impact (Warren et al., 13 Feb 2025).

Known limitations include:

• Extraction Noise and Ontological Gaps: Event extraction (e.g., REBEL) and evidence retrieval may introduce abstraction or coverage errors.
• Lack of fine-grained causality and counterfactual reasoning: Most systems still lack full support for “what-if” or causal explanations.
• Incomplete multi-modal support: Visual/chart/social inputs require tailored explainers and fusion schemes (Akhtar et al., 2023).
• Subjectivity in explanation evaluation: Convincingness and overall coherence remain challenging to quantify automatically (Feher et al., 2024).
• Scalability and throughput: Over-reliance on external LLMs, live search APIs, or complex user feedback pipelines may constrain operational deployment (Althabiti et al., 2024, Zhang et al., 2021).

6. Future Directions and Open Problems

Pressing research goals for EAFC include:

• Integrating structured causal reasoning and counterfactual generation to fully address causality and support reasoning over nested modalities and negations (Rebboud et al., 15 Dec 2025).
• Extending explainability to rich multi-modal domains (images, tables, charts) via unified vision–LLMs and multi-hop reasoning (Akhtar et al., 2023, Mahmood et al., 2024).
• Standardized, scalable evaluation datasets for explanations, covering actionability, faithfulness, and bias (Eldifrawi et al., 7 Apr 2025, Kotonya et al., 2020).
• Advancing user-adaptive and dialogic explanation frameworks (REFLEX), leveraging internal model representations, contrastive activation steering, and chain-of-thought guidance (Kong et al., 25 Nov 2025).
• Developing interactive systems supporting continual human correction, explanation refinement, and life-long learning (Zhang et al., 2021, Warren et al., 13 Feb 2025).

System designers must balance trade-offs among explanation completeness, clarity, faithfulness to internal computation, user auditability, and efficiency, with dynamic adaptation to both domain-specific and real-world user workflows.

---

In summary, explainable automated fact-checking encompasses a spectrum of paradigms (rule-based, neural, causal, multi-modal, multi-hop, and interactive) unified by their commitment to not only classifying claims but also producing transparent, actionable, and technically rigorous justifications—addressing the epistemic, legal, and practical demands of trustworthy, scalable misinformation mitigation (Rebboud et al., 15 Dec 2025, Lourenço et al., 11 Aug 2025, Kotonya et al., 2020, Eldifrawi et al., 2024, Eldifrawi et al., 7 Apr 2025, Rani et al., 2023, Yang et al., 2021, Akhtar et al., 2023, Kong et al., 25 Nov 2025).