Claim–Evidence Graph

• Claim–evidence graphs are structured representations linking claims with supporting evidence using graph-theoretic formalisms.
• They employ diverse construction methods, including sentence-level, semantic, token co-occurrence, and triplet/entity-relation approaches for detailed fact-checking.
• These graphs enable explainable decision-making and robust evidence aggregation, driving state-of-the-art performance in misinformation detection.

A claim–evidence graph is a structured representation that models the relationships between claims (statements to be verified) and their associated evidential support, leveraging graph-theoretic formalisms to enable fine-grained multi-evidence reasoning, rigorous aggregation, and explainable fact verification in both unimodal (textual) and multimodal settings. Claim–evidence graphs can be instantiated at various granularities: as sentence-level, semantic-unit–level, or entity–relation graphs, and are central to state-of-the-art fact-checking and misinformation detection systems across several recent research paradigms.

1. Formal Definitions and Graph Construction Paradigms

Claim–evidence graphs are typically instantiated in several structurally distinct forms, each grounded in the specific reasoning needs and data sources of the fact verification task.

1. Sentence/evidence node graphs: Each node represents an evidence sentence (retrieved or selected via a retriever), and edges encode hypothesized dependencies or enable message-passing for information flow. Notably, the claim is not an explicit node but is “baked” into evidence node features via joint encoding (e.g., BERT over $(e_i, c)$). Fully connected evidence graphs with self-loops are constructed, as in GEAR (Zhou et al., 2019).
2. Semantic-level graphs: Nodes are semantic units extracted via semantic role labeling (predicate–argument structures) or span-based heuristics in both claims and evidence. Edges are of two types: intra-tuple (within SRL tuples) and cross-tuple (based on semantic or lexical overlap). These graphs capture fine-grained semantic dependencies and enable reasoning about multi-hop relations (Zhong et al., 2019).
3. Word/Token co-occurrence graphs: Each unique token in claim or evidence text is a node, with undirected edges linking tokens that co-occur within a sliding window. Graph-structure learning modules prune redundant nodes to minimize information overload and noise propagation (Xu et al., 2022, Wu et al., 2022). These graphs support evidence-aware fake news detection by modeling long-distance semantic dependencies via GGNN propagation and redundancy-aware pruning.
4. Triplet/entity-relation graphs: Both claims and evidence are decomposed into sets of atomic subject–predicate–object triplets. The claim graph is built from LLM or rule-based triplet extraction, with entities as nodes and predicates as labeled edges. Evidence graphs represent known entities and their relations, usually restricted by entity disambiguation. Verification is then formulated as mapping (via bijection or soft matching) from claim triplets to evidence triplets, resolving ambiguities and supporting fine-grained logical verification (Huang et al., 10 Mar 2025, Pham et al., 29 May 2025).
5. Multimodal and heterogeneous graphs: In tasks involving multiple modalities (e.g., image–caption verification), claim–evidence graphs have multiple node and edge types (e.g., text nodes, image nodes, and their cross-modal links), with explicit encoding of entity/event nodes, attribute relations (PERFORMS, LOCATED_IN, etc.), and multimodal transformer-GNN architectures for cross-modal evidence fusion (Duwal et al., 23 May 2025, Zhang et al., 10 Feb 2026).
6. KG-anchored and attribution graphs: For explainable claim-level analysis, bipartite graphs link claim spans to KG triplets, with edges annotated by attribution scores measuring the degree of semantic/entity support or contradiction. Node and edge scores aggregate to global attribution metrics (Dammu et al., 2024).

2. Evidence Aggregation, Message Passing, and Relational Reasoning

Claim–evidence graphs provide the substrate for sophisticated information propagation, relational reasoning, and multi-evidence fusion, generally realized via GNN variants, attention pooling, and prompt-based fusion mechanisms.

• Graph neural propagation: GAT, GGNN, and TransformerConv layers propagate information among evidence sentence nodes or semantic units. Attention coefficients (learned via MLP or parameterized inner-product) weight information transfer to focus propagation on salient nodes/edges and mitigate noisy or adversarial signals (Zhou et al., 2019, Wu et al., 2022).
• Redundancy-aware structure learning: Graph-structure learning modules iteratively prune redundant or spurious nodes/edges. Node redundancy scores (estimated by downstream classifiers or GNNs) control edge/row deletions in the adjacency matrix, and only salient nodes remain for further propagation (Xu et al., 2022, Wu et al., 2022).
• Claim–evidence interaction: Cross-graph attention mechanisms align claim graph nodes to evidence nodes, allowing claim units to pull relevant evidence pools via attention-weighted summations. Dual-pool aggregators (word/node- and document-level) enable hierarchical evidence fusion (Zhong et al., 2019, Xu et al., 2022).
• Prompt augmentation and context/reference integration: In multi-layered graphs, as in CORRECT (Zhang et al., 9 Feb 2025), each evidence node is contextually updated via GNN-style message passing within and across layers (evidence, context, reference). These graph-derived features condition subsequent prompt vectors or token sequences, which are then used to steer transformers or LLMs for verdict prediction.
• Noise suppression and confidence gating: Node-masking or confidence-scoring schemes pre-attend to relevance and gate the influence of uninformative or noisy nodes in propagation. For example, CO-GAT (Lan et al., 2024) softly projects low-confidence evidence nodes toward a blank-claim vector, suppressing noise transfer during GAT updates.

3. Verification, Decision Fusion, and Explainability

Claim–evidence graphs structure the aggregation of reasoning results across multiple evidential units and enable interpretable decision-making.

• Attention-based evidence aggregation: Final node features are typically collapsed (via claim-aware attention pooling, mean/max pool, or global attention heads) into a single representation $o$ that conditions the downstream classifier (Zhou et al., 2019).
• Joint claim+evidence modeling: Architectures such as CORRECT and GEAR inject claim semantics into all processing stages—first via conditioning evidence node encodings and later by using standalone claim vectors in pooling/attention aggregators.
• Graph-guided sequential verification: For triplet-based graphs, planned, dependency-ordered checking of sub-claims/triplets (e.g., grounded triples before resolving unknowns) allows structured control over multi-hop verification and referential ambiguity (Huang et al., 10 Mar 2025, Pham et al., 29 May 2025).
• Scoring, attribution, and rationale generation: Claim–evidence edges are annotated with match scores (e.g., Triplet Match Score in ClaimVer, which symmetrically combines entity overlap and semantic similarity (Dammu et al., 2024)). Node and global attribution scores quantify support, contradiction, or neutrality. In MEVER (Zhang et al., 10 Feb 2026), a fusion-in-decoder approach generates natural-language explanations for verification labels, with a consistency regularizer ensuring faithfulness to model evidence.
• Logic-level aggregation: Some frameworks enforce strict logical AND across triplet verifications (VeGraph (Pham et al., 29 May 2025)), while others employ weighted, soft fusion (e.g., attention or mean pooling). Such logic is crucial for both explainability and precision in complex, multi-hop reasoning contexts.

4. Algorithmic Pipelines and System Architecture

Implementations of claim–evidence graph–based fact verification are typically decomposed into multi-stage pipelines:

System	Graph Construction	Feature Propagation	Evidence Aggregation	Final Prediction
GEAR (Zhou et al., 2019)	Sent/node graph, BERT $(e_i, c)$	ERNet (GAT attention)	Pooling or attention	MLP + softmax
GET/GETRAL (Xu et al., 2022, Wu et al., 2022)	Word/token co-occurrence	GGNN, redundancy pruning	Node/doc-level attention	MLP + adversarial/contrastive
DREAM (Zhong et al., 2019)	Semantic-unit (SRL) graphs	GCN, cross-graph attention	Joint pooling	MLP + softmax
CORRECT (Zhang et al., 9 Feb 2025)	Triple-layer node types	GNN in transformer steps	Mean across evidences	Prompt-based transformer obj.
GraphFC (Huang et al., 10 Mar 2025)	Triplet graph (claim, evidence)	Graph-guided planning/checking	Sequential triplet verif	Boolean (AND/OR) fusion
CO-GAT (Lan et al., 2024)	Claim–evidence pair nodes	Confidence masking + multi-head GAT	Node-level attention	MLP + softmax
VeGraph (Pham et al., 29 May 2025)	LLM-extracted triplet graph	Iterative disambiguation	Subclaim-level checking	AND aggregation
MEVER (Zhang et al., 10 Feb 2026)	Heterogeneous text/image	Cross-modal GNN, attention	Token- and doc-level	MLP + joint seq2seq decoder

Typical end-to-end workflows encompass retrieval of candidate documents, selection of salient sentences, graph construction by evidence and/or semantic decomposition, iterative GNN propagation and pooling, and final prediction and explanation.

5. Extensions: Multimodal, Causal, and Knowledge-based Variants

Recent work extends the claim–evidence graph paradigm to admit more diverse data modalities and reasoning paradigms.

• Multimodal extension: In MEVER (Zhang et al., 10 Feb 2026) and EGMMG (Duwal et al., 23 May 2025), heterogeneous graphs connect textual and visual evidence, with joint image–text node representations. Graph neural reasoning in these settings supports cross-modal claim verification and generates faithful explanations.
• Causal reasoning and confounder control: MuPlon (Guo et al., 30 Sep 2025) treats the C-E graph as an undirected, fully-connected node set. Dual-path causal interventions address both data noise (back-door path: Bayesian node probability weighting, node generation) and data bias (front-door path: Markov path extraction, counterfactual bias removal via confusion dictionaries and attention pooling).
• Knowledge graph integration: KG-CRAFT (Lourenço et al., 27 Jan 2026) and ClaimPKG (Pham et al., 28 May 2025) explicitly link claims and evidence to external knowledge graphs. Contrastive question generation based on KG triples enhances the focus of LLM-based verification, and pseudo-subgraph generation provides lightweight interfaces for scalable graph retrieval and LLM integration. Edge annotation, entity typing, and confidence propagation are supported via LLM-in-the-loop extraction and ranking.
• Explainability and score attribution: ClaimVer (Dammu et al., 2024) introduces a layered scoring mechanism (claim score, triplet match score, KG attribution score) with a bipartite claim–evidence edge graph that allows fine-grained, explainable evidence attribution and model introspection.

6. Empirical Validation and Impact on Multi-hop Fact Verification

Extensive evaluation on benchmark datasets (FEVER, FEVEROUS, HOVER, FactKG, RAWFC, LIAR-RAW, PolitiFact) demonstrates:

• Superior multi-evidence reasoning: Claim–evidence graphs significantly outperform sequential and flat models in multi-hop scenarios, yielding state-of-the-art FEVER scores (GEAR: 67.10% (Zhou et al., 2019), DREAM: 70.60% (Zhong et al., 2019)), and high macro-F1 gains on open-domain, adversarial, and domain-specific tasks (Huang et al., 10 Mar 2025, Guo et al., 30 Sep 2025).
• Noise and bias mitigation: Methods such as redundancy pruning, node confidence weighting, and dual causal adjustment suppress spurious signals and over-smoothing, improving reliability, robustness, and generalization (Wu et al., 2022, Lan et al., 2024, Guo et al., 30 Sep 2025).
• Explainability and coverage: Graph-based approaches natively support evidence attribution, subclaim-level rationales, and local/global score explanations (Dammu et al., 2024, Zhang et al., 10 Feb 2026).
• Cross-modal and reference awareness: Multimodal extensions demonstrate disproportionate gains against LLM and LVLM baselines in mismatched or out-of-context detection tasks (Duwal et al., 23 May 2025). Hybrid graph–prompt systems yield robust performance even in low-data, few-shot regimes (Zhang et al., 9 Feb 2025).
• Generalization to complex and ambiguous claims: Graph-centric decomposition, entity disambiguation, and graph-guided planning resolve referential ambiguities and cope with latent entity mentions better than flat LLM-based pipelines (Pham et al., 29 May 2025, Huang et al., 10 Mar 2025).

7. Research Directions and Open Problems

Despite substantial advances, several open fronts remain:

• Combinatorial graph construction efficiency: Large-scale, online graph construction (e.g., whole-corpus entity co-mention graphs (Mongiovì et al., 2021)) and fine-grained triplet extraction at scale remain computationally intensive and require further architecture optimization.
• Integration with general-purpose and task-specific LLMs: Careful pipelining and interface design are required to combine lightweight specialized LLM modules (pseudo-subgraph generation) with general-purpose LLM reasoning engines (Pham et al., 28 May 2025).
• Adaptive graph structure learning: Continuous adaptation of graph topology—guided by reinforcement signals or learned relevance weights—offers paths to further mitigate noise, handle weak evidence, and scale to even more ambiguous or adversarial inputs.
• Multilingual and multimodal unification: Extending the graph-based paradigm across languages and imaging modalities while maintaining robust reasoning and explainability remains an active area of research.
• Human-in-the-loop and probabilistic modeling: Graph-based explanations and attributions provide opportunities for downstream integration with human fact-checkers, iterative feedback loops, and uncertainty quantification in verification decisions (Lourenço et al., 27 Jan 2026).

Claim–evidence graphs have become a foundational abstraction for multi-evidence fact verification, enabling structured, interpretable, and high-fidelity reasoning across text, knowledge graphs, and multimodal domains, and continue to underpin empirical advances in automated fact-checking and misinformation detection.