Graph-based Evidence Aggregating and Reasoning (GEAR): A Graph Neural Network Approach to Fact Verification

With the rapid maturation of natural language processing (NLP) technologies, many language tasks have shifted from manual to automated processing, but advanced applications like knowledge‑graph construction and open‑domain question answering require highly accurate and trustworthy textual evidence, making fact verification increasingly important.

Fact verification involves judging whether a claim is supported, contradicted, or lacks sufficient information based on multiple pieces of evidence, a task that is challenging because it often requires synthesizing several evidence fragments to form new logical connections.

In practice, evidence extraction modules can introduce noisy or irrelevant evidence, further complicating the verification process, as illustrated by the provided examples.

The research team from Tsinghua University proposes the GEAR model (Graph‑based Evidence Aggregating and Reasoning), which first builds a fully connected evidence graph, encourages information propagation among evidence nodes, and then uses graph attention mechanisms to aggregate evidence fragments before classifying the claim’s support status.

All input sentences are first encoded with a strong pre‑trained language model such as BERT, providing rich semantic representations that the graph‑based reasoning component can exploit.

Evaluated on the FEVER fact‑extraction and verification dataset, GEAR achieves a 3‑4% accuracy improvement over previous methods and shows superior performance on samples requiring multi‑evidence reasoning.

Beyond fact‑checking, GEAR can be applied to detect misinformation on social media, enhance knowledge‑graph completion, improve open‑domain QA, and even assess user behavior credibility for advertising platforms.

The authors also authored a comprehensive review of graph neural networks, outlining existing models, applications, and open research questions.

This presentation is part of the Wiztalk Tencent series, which aims to disseminate cutting‑edge research topics to a broader audience.