CosG: A Graph-Based Contrastive Learning Method for Fact Verification.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Information Retrieval

Background:

• Fact verification relies on assessing claim authenticity using evidence from sources like Wikipedia.
• Current methods often struggle to differentiate semantically similar claims and evidence with opposing authenticity labels.
• Graph neural networks in existing models face limitations like over-smoothing, hindering performance.

Purpose of the Study:

• To develop a fact verification method that effectively distinguishes between claims and evidence with different authenticity labels.
• To address the limitations of graph neural networks, such as over-smoothing and loss of unique node features.
• To enhance the robustness of fact verification models, particularly in low-resource scenarios.

Main Methods:

• Proposed CosG, a graph-based contrastive learning method for fact verification.
• Introduced a label-supervised contrastive task to learn discriminative representations for claim-evidence pairs.
• Incorporated an unsupervised graph-contrast task to mitigate unique node feature loss during graph propagation.

Main Results:

• CosG demonstrated superior performance compared to existing baselines on the FEVER dataset.
• The method showed particular effectiveness in verifying claims requiring multiple pieces of evidence.
• CosG exhibited enhanced model robustness in low-resource fact verification scenarios.

Conclusions:

• CosG offers a significant advancement in fact verification by improving the discriminative power of models.
• The proposed contrastive learning approach effectively handles complex claims and enhances model robustness.
• This method shows promise for improving the reliability of automated fact-checking systems.