Heuristic Heterogeneous Graph Reasoning Networks for Fact Verification.

Area of Science:

• Artificial Intelligence
• Data Science
• Computer Science

Background:

• Existing table-based fact verification methods analyze linguistic and logical evidence separately.
• Limited interaction between these evidence types hinders the extraction of valuable consistency features.

Purpose of the Study:

• To propose heuristic heterogeneous graph reasoning networks (H2GRN) for enhanced table-based fact verification.
• To strengthen associations between linguistic and logical evidence for improved consistency feature extraction.

Main Methods:

• Constructed a heuristic heterogeneous graph using claim semantics to guide program-table subgraph connections and program logic to expand claim-table subgraph connectivity.
• Designed multi-view reasoning networks, including local-view multi-hop knowledge reasoning (MKR) and global-view dual-attention networks (DAN).
• Employed a consistency fusion layer to reconcile differing evidence and capture shared consistent evidence.

Main Results:

• H2GRN effectively captures shared consistent evidence by strengthening associations between linguistic and logical evidence.
• The proposed methods enhance the connectivity and interaction between claim-table and program-table subgraphs.
• Experiments on TABFACT and FEVEROUS datasets demonstrate the effectiveness of H2GRN.

Conclusions:

• H2GRN offers a novel approach to table-based fact verification by integrating linguistic and logical evidence through graph reasoning.
• The method improves the ability to identify consistent evidence crucial for accurate claim verification.
• The findings highlight the potential of heuristic graph construction and multi-view reasoning for complex data verification tasks.