Research map / Graph learning / Knowledge and temporal graphs
Heterogeneous Graph Reasoning for Fact Checking over Texts and Tables
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HeterFC, a word-level heterogeneous evidence graph with R-GCN propagation, attention-based claim-evidence aggregation, and an auxiliary evidence-relevance loss for fact checking over text and tables on FEVEROUS.
Problem
Veracity reasoning for fact checking over mixed unstructured text and structured table evidence, where prior transformer concatenation and sentence-level homogeneous graphs miss fine-grained entity-level and table-structure relations and are hurt by noisy retrieved evidence.
Approach
Builds a word-level heterogeneous evidence graph with three edge types (inter-evidence edges between shared keywords, intra-sentence and intra-table sliding-window edges), propagates with R-GCN, reads out per-evidence embeddings, aggregates them with claim-guided attention trained by an auxiliary evidence-relevance loss, and fuses with a RoBERTa claim-evidence encoding.
Key results
- On the FEVEROUS blind test set, HeterFC reaches FEVEROUS score 0.3476 and label accuracy 0.6329, versus 0.3397 / 0.6321 for DCUF.
- Replacing word-level graph with sentence-level graph (HeterFC_sent) causes a significant drop, while token-level graph is slightly worse than word-level.
- Dev-set results with a simpler retriever show HeterFC still leads (FEVEROUS 0.2427, label accuracy 0.6977).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| word-level heterogeneous evidence graph (HeterFC) vs. baselines | FEVEROUS test veracity reasoning · FEVEROUS test | Feverous score: 0.3476 vs 0.3397 (DCUF best baseline) | Helped |
| replace attention-based claim-evidence interaction with mean pooling (ablation) | HeterFC · FEVEROUS dev | Feverous score: 0.3613 vs 0.3714 | Hurt |
| replace attention-based claim-evidence interaction with max pooling (ablation) | HeterFC · FEVEROUS dev | Feverous score: 0.3620 vs 0.3714 | Hurt |
| remove assisted veracity representation o_t (graph-only) (ablation) | HeterFC · FEVEROUS dev | Feverous score: 0.3640 vs 0.3714 | Hurt |
| remove assisted evidence loss Loss_e (ablation) | HeterFC · FEVEROUS dev | Feverous score: 0.3641 vs 0.3714 | Hurt |
| word-level nodes instead of sentence-level graph nodes (ablation) | HeterFC · FEVEROUS dev | Feverous score and label accuracy: sentence-level variant significantly lags (values only in figure) | Helped |
| token-level nodes (no subword mean pooling) instead of word-level (ablation) | HeterFC · FEVEROUS dev | Feverous score: HeterFC better Feverous score, slightly better label accuracy (values only in figure) | Hurt |
| remove heterogeneous relations (homogeneous edges) (ablation) | HeterFC · FEVEROUS dev | Feverous score: marked drop (values only in figure) | Hurt |
| remove local/global edge design (fully-connected graph, heterogeneous relations kept) (ablation) | HeterFC · FEVEROUS dev | Feverous score: marked drop (values only in figure) | Hurt |
| remove all special edge designs (fully-connected, homogeneous) (ablation) | HeterFC · FEVEROUS dev | Feverous score: weakest variant (values only in figure) | Hurt |
| increase R-GCN layers beyond k=2 (ablation) | HeterFC · FEVEROUS dev | Feverous score and label accuracy: metrics drop beyond k=2 (values in figure) | Hurt |
| use RoBERTa-Pair mean vs. max vs. Concat transformer baselines | transformer baselines · FEVEROUS test | Feverous score: 0.3231/0.3347/0.3344 | No clear effect |
Surprising findings
- Token-level nodes underperform word-level nodes because shared subwords (e.g., 'ing') create spurious inter-evidence edges.
- Homogeneous fully connected graph variants (GEAR, KGAT) perform worse than simple RoBERTa-based transformer baselines on this task.
Experimental setup
Datasets: FEVEROUS
Benchmarks: FEVEROUS
Metrics: FEVEROUS score label accuracy
Baselines: RoBERTa-Pair (mean) RoBERTa-Pair (max) RoBERTa-Concat DCUF GEAR KGAT
Builds on: R-GCN GEAR KGAT DCUF RoBERTa
Closest papers by meaning
- Beyond Text: Fine-Grained Multi-Modal Fact Verification with Hypergraph Transformers (AAAI 2025 · similarity 0.89)
- Enhancing Multi-Hop Fact Verification with Structured Knowledge-Augmented Large Language Models (AAAI 2025 · similarity 0.87)
- VeriGraph: Towards Verifiable Data-Analytic Agents (NeurIPS 2026 · similarity 0.86)
- MRR-FV: Unlocking Complex Fact Verification with Multi-Hop Retrieval and Reasoning (AAAI 2025 · similarity 0.86)
- Conformal Language Model Reasoning with Coherent Factuality (ICLR 2025 · similarity 0.85)
- TopoGraphRAG-Bench: Evaluating Multimodal GraphRAG on Layout-Grounded Evidence Reasoning (NeurIPS 2026 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.