Research map / Probabilistic and scientific modeling / Treatment effects
Causal Walk: Debiasing Multi-Hop Fact Verification with Front-Door Adjustment
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 Causal Walk, which debiases multi-hop fact verification via front-door adjustment with reasoning paths (random-walk sampled) as mediator, plus NWGM approximation, and builds GPT-4-extended Hard/Symmetric PolitiHop datasets.
Problem
Multi-hop fact verification models rely on spurious annotation biases hidden across multiple evidence hops, causing performance drops on unbiased or adversarial data; existing causal debiasing methods only handle single-hop shallow biases.
Approach
Introduces the reasoning path over the claim-evidence graph as a mediator in a structural causal model, enabling front-door adjustment; estimates P(r|do(G)) via random-walk path sampling on a GNN and P(L|do(r)) via LSTM path encoding with NWGM approximation over a label-indexed confounder dictionary.
Key results
- Causal Walk reaches 80.00% on PolitiHop and 79.00% on Adversarial PolitiHop, versus 77.00% for KGAT and 78.00% for CICR-graph/CLEVER-graph (best baselines).
- On Adversarial FEVER-MH Causal Walk gets 67.08% vs 64.85% for KGAT.
- On Symmetric PolitiHop Causal Walk scores 57.02%, above GEAR's 51.17% and CLEVER-graph's 52.05%, though it is below GEAR on PolitiHop-adv (88.30 vs 89.47).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove causal intervention (ablation) | Causal Walk · PolitiHop | accuracy: 80.00 -> 77.00 | Hurt |
| remove causal intervention (ablation) | Causal Walk · Adversarial PolitiHop | accuracy: 79.00 -> 78.00 | Hurt |
| remove causal intervention (ablation) | Causal Walk · Hard PolitiHop | accuracy: 79.00 -> 77.00 | Hurt |
| remove causal intervention (ablation) | Causal Walk · FEVER | accuracy: 90.19 -> 87.86 | Hurt |
| remove causal intervention (ablation) | Causal Walk · FEVER-MH | accuracy: 92.88 -> 90.69 | Hurt |
| remove causal intervention (ablation) | Causal Walk · Adversarial FEVER | accuracy: 62.13 -> 60.86 | Hurt |
| remove causal intervention (ablation) | Causal Walk · Adversarial FEVER-MH | accuracy: 67.08 -> 65.35 | Hurt |
| add evidence-label supervision for the transition probability matrix (ablation) | Causal Walk · PolitiHop | accuracy: 80.00 -> 78.00 | Hurt |
| add evidence-label supervision for the transition probability matrix (ablation) | Causal Walk · Adversarial PolitiHop | accuracy: 79.00 -> 77.00 | Hurt |
| add evidence-label supervision for the transition probability matrix (ablation) | Causal Walk · Hard PolitiHop | accuracy: 79.00 -> 77.00 | Hurt |
| add evidence-label supervision for the transition probability matrix (ablation) | Causal Walk · FEVER | accuracy: 90.19 -> 89.48 | Hurt |
| add evidence-label supervision for the transition probability matrix (ablation) | Causal Walk · FEVER-MH | accuracy: 92.88 -> 91.87 | Hurt |
Surprising findings
- Using ground-truth evidence labels to supervise the path transition matrix decreased performance, especially on adversarial data, which the authors attribute to sparse transition matrices.
Limitations
- Symmetric results are modest (e.g., 54.09 on Symmetric-ShareEvi), and Causal Walk is not best on PolitiHop-adv.
- Hard and Symmetric PolitiHop datasets are built from GPT-4 generations, whose quality is not systematically validated here.
- Evaluated only on BERT-base with small batch size and only FEVER/PolitiHop-based datasets.
Experimental setup
Datasets: FEVER PolitiHop
Benchmarks: PolitiHop Adversarial PolitiHop Hard PolitiHop FEVER FEVER-MH Adversarial FEVER Adversarial FEVER-MH Symmetric PolitiHop
Metrics: accuracy
Baselines: BERT-Concat GEAR KGAT Transformer-XH CICR CLEVER CICR-graph CLEVER-graph
Builds on: front-door adjustment NWGM approximation CICR CLEVER KGAT
Closest papers by meaning
- MRR-FV: Unlocking Complex Fact Verification with Multi-Hop Retrieval and Reasoning (AAAI 2025 · similarity 0.87)
- Grounding Multi-Hop Reasoning in Structural Causal Models via Group Relative Policy Optimization (ICML 2026 · similarity 0.86)
- Enhancing Multi-Hop Fact Verification with Structured Knowledge-Augmented Large Language Models (AAAI 2025 · similarity 0.84)
- Beyond Text: Fine-Grained Multi-Modal Fact Verification with Hypergraph Transformers (AAAI 2025 · similarity 0.84)
- BiDeV: Bilateral Defusing Verification for Complex Claim Fact-Checking (AAAI 2025 · similarity 0.83)
- Prompting Fairness: Integrating Causality to Debias Large Language Models (ICLR 2025 · similarity 0.83)
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.