Denoising Rationalization for Multi-hop Fact Verification via Multi-granular Explainer

Jiasheng Si, Yingjie Zhu, Wenpeng Lu, Deyu ZhouEMNLP 2024 · DOI

Abstract

Deep learning models for multi-hop fact verification often suffer from "noise rationalization," where they identify irrelevant but correlated tokens as evidence. We present CURE (Consistent mUlti-granular Rationale Extraction), a framework that filters out noise by enforcing collaboration between token-level and sentence-level explainers. By ensuring consistency across granularities and intervening in the model's reasoning process, CURE produces rationales that are both faithful to the model and logically sound for human users.

Introduction

Multi-hop fact verification requires connecting disparate pieces of evidence to verify a claim. While modern models achieve high accuracy, their explanations (rationales) often include "noise evidence"—sentences containing words correlated with the label but logically irrelevant to the claim. This "Right Prediction for Wrong Reasons" phenomenon undermines trust. Traditional frameworks often fail to distinguish between true logical chains and high-entropy noise that happens to mimic evidence.

Comparison of traditional rationalization frameworks versus the proposed denoising approach.

Methodology: The CURE Framework

CURE treats rationalization as a collaborative effort between two parallel explainers: one focused on individual tokens and another on entire sentences.

The architecture of the CURE framework, showing the interaction between token and sentence explainers and the veracity model.

The framework operates on a pre-trained (and frozen) veracity prediction model. Key technical innovations include:

  • Multi-granular Collaboration: The final rationale is derived by intersecting the outputs of the token-level and sentence-level explainers. If a token is flagged as important but resides in a sentence deemed "noise," it is discarded.
  • Diagnostic Properties: CURE optimizes for Fidelity (ensuring the rationale is sufficient for the model), Consistency (aligning importance scores via Jensen-Shannon Divergence), and Salience (focusing on "true" evidence using human-annotated labels and Layered Integrated Gradients).
  • Capsule Network Intervention: The sentence-level mask intervenes directly in the veracity model's dynamic routing, suppressing the influence of noise sentences during the reasoning process.
  • Noiselessness Metric: A new evaluation metric designed to measure the "purity" of a rationale—specifically, how much of the extracted token spans come from true evidence versus noise evidence.

Datasets and Evaluation

The model was evaluated on three challenging multi-hop fact-verification benchmarks: HoVer (Wikipedia multi-hop reasoning), LIAR-PLUS (human-written justifications), and PolitiHop (political claims with annotated noise).

CURE outperformed 12 baseline methods (including LIME, SHAP, and VMASK) across all datasets:

  • Superior Purity: On the HoVer dataset, CURE achieved a Noiselessness score of 0.9522, compared to 0.6474 for the best-performing baseline (DiffMask). This indicates that CURE's rationales are almost entirely free of irrelevant noise.
  • Faithfulness: CURE achieved the lowest Fidelity scores, indicating the rationale is highly sufficient to explain the prediction.
  • Human Alignment: In manual evaluations, human annotators rated CURE significantly higher than baselines in terms of Correctness and Faithfulness, noting its ability to retain document titles—the "key cues" used to link multiple pieces of evidence.

Discussion and Conclusion

The core insight of CURE is that noise often exists at one granularity but not another. A "noise" sentence might contain a correlated token, but the sentence itself lacks logical flow. By requiring consistency across levels, the model filters out these spurious correlations. While CURE currently relies on supervised sentence-level labels, future work will explore semi-supervised or unsupervised variants and counterfactual explanations to further enhance transparency in multi-hop reasoning.

Summary updated

BibTeX
@inproceedings{si2024denoising,
  author = {Jiasheng Si and Yingjie Zhu and Wenpeng Lu and Deyu Zhou},
  title = {{Denoising Rationalization for Multi-hop Fact Verification via Multi-granular Explainer}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2024},
  doi = {10.18653/v1/2024.findings-emnlp.736},
}
Figure