A syntactic evidence network model for fact verification.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Fact verification is crucial for assessing claim authenticity using evidence.
• Current deep learning methods often struggle with irrelevant word attention in fact verification.
• Existing models lack specific constraints on important claim and evidence words.

Purpose of the Study:

• To propose a novel Syntactic Evidence Network (SENet) model for enhanced fact verification.
• To address limitations in current attention mechanisms by incorporating syntactic information.
• To improve the accuracy and performance of automated fact-checking systems.

Main Methods:

• Developed the SENet model integrating entity keywords, syntactic information, and sentence attention.
• Utilized a pre-trained syntactic dependency parser to extract sentence structures.
• Incorporated syntactic information into the attention mechanism for language-driven word representation.

Main Results:

• Achieved 78.69% Label Accuracy and 75.63% FEVER Score on the FEVER dataset.
• Obtained 65.0% precision and 61.2% macro F1 on the UKP Snopes dataset.
• Demonstrated superior performance compared to baseline models in fact verification tasks.

Conclusions:

• The SENet model significantly enhances fact verification accuracy.
• Integrating syntactic information and targeted attention improves semantic representation.
• SENet achieves state-of-the-art performance, outperforming existing fact-checking methods.