TL;DR
Deep learning models for multi-hop fact verification often operate as black boxes, making it hard to see how they connect multiple pieces of evidence. This paper introduces CURE (Consistent mUlti-granular Rationale Extraction), a framework that simultaneously extracts sentence-level and token-level explanations. By enforcing consistency between these two granularities, CURE delivers highly precise, faithful, and noise-free explanations without compromising the model's predictive accuracy.
What Problem Does This Paper Address?
To verify complex claims, fact-checking systems must perform "multi-hop" reasoning—linking multiple disparate pieces of evidence together. While modern neural networks excel at this, explaining why a model reached a specific conclusion remains a challenge.
A popular explainability technique is erasure search: finding the smallest subset of the input text (the "rationale") that can be kept while still preserving the model's original prediction. However, existing erasure methods only look at a single granularity:
- Sentence-level rationales are too coarse, often retaining irrelevant words and noise evidence alongside the facts.
- Token-level rationales lack structural context, frequently selecting scattered words from completely irrelevant sentences due to spurious correlations.
This single-granularity focus leads to redundant or contradictory explanations. This paper addresses this gap by asking: How can we extract token-level and sentence-level rationales simultaneously so that they complement and validate each other?
Methods
The authors propose CURE, a post-hoc explainability framework. It takes a frozen, pretrained multi-hop fact verification model (specifically, a Transformer-XH architecture paired with a Capsule Network) and trains two lightweight, parameterized explainer networks on top of it.
The core innovation lies in how these explainers are trained using differentiable masking (via the Hard Concrete reparameterization trick) guided by three diagnostic properties:
- Fidelity: Ensures the extracted rationales are sufficient on their own to preserve the original model's prediction. This is enforced by minimizing the Euclidean distance between the prediction logits of the original text and the masked text.
- Consistency: Uses symmetric Jensen-Shannon Divergence to align the token-level and sentence-level explainers. This forces the token explainer to select words primarily from the sentences that the sentence explainer deemed important, preventing it from picking random tokens from noise sentences.
- Salience: Uses supervised signals as prior knowledge to guide the search space. It leverages human-annotated sentence labels and computes token-level pseudo-labels using Layered Integrated Gradients.
Datasets
The framework was evaluated on three challenging multi-hop fact verification benchmarks:
- HoVer: A dataset specifically designed for many-hop fact extraction and claim verification, where each claim is paired with 5 retrieved text snippets.
- LIAR-PLUS: A benchmark containing political claims, each associated with 10 pieces of evidence.
- PolitiHop: A dataset focusing on political fact-checking with 5 pieces of evidence per claim.
All three datasets include human annotations for sentence-level rationales mixed with distracting noise evidence.
Key Insights and Ideas
- Granular Synergy: Training token and sentence explainers together acts as a powerful mutual regularizer. Sentence-level context prevents the token explainer from hallucinating importance, while fine-grained token importance helps sharpen sentence selection.
- Pruning Spurious Correlations: Without consistency constraints, token-level explainers suffer from "under-aggressive pruning"—they fail to ignore tokens in completely irrelevant sentences. Intersecting the two granularities effectively shields the explanation from data artifacts and noise.
- Anchoring on Titles: Qualitatively, CURE demonstrates a strong tendency to retain document titles. The model recognizes that titles serve as critical semantic "bridges" necessary for linking multi-hop evidence across different documents.
Key Results
CURE significantly outperformed classical post-hoc explainers (like LIME, SHAP, and DiffMask) across several metrics:
- Superior Consistency: CURE achieved a consistency score of 0.9522 on the HoVer dataset, whereas baseline methods often hovered near zero or even showed negative consistency (meaning they picked tokens from sentences they claimed were irrelevant).
- High Faithfulness: CURE achieved the lowest Fidelity loss across all three datasets (e.g., 0.2287 on HoVer), proving that its concise snippets truly reflect the internal reasoning of the verification model.
- Human Approval: In a blind evaluation by human annotators rating explanations on a 1–5 scale, CURE scored the highest on average across Correctness, Faithfulness, and Non-redundancy.
Limitations and Future Work
Despite its success, CURE has a few limitations:
- Supervision Dependency: The "Salience" property relies on human-annotated sentence-level rationale labels during training. This makes it difficult to deploy CURE in fully unsupervised scenarios where rationale labels are missing.
- Sensitivity to Sparsity: The framework uses an $L_0$ regularization loss to keep explanations concise. However, overemphasizing sparsity can occasionally strip away context that is minorly helpful to the verification model, making the framework sensitive to hyperparameter tuning.
For future work, the authors plan to extend this multi-granular approach to generate counterfactual explanations (i.e., demonstrating how changing a specific fact would alter the model's verdict).
Summary updated