Graph-based Evidence Aggregating and Reasoning (GEAR) Model for Fact Verification in NLP

As natural language processing (NLP) matures, tasks are moving from simple text handling to more advanced applications such as knowledge‑graph construction and open‑domain question answering, which require highly accurate corpora.

Fact verification—determining whether a claim is supported, refuted, or lacks sufficient evidence—has become crucial both for downstream NLP tasks and for combating misinformation in today’s fast‑moving media environment.

The challenge lies in the need to combine multiple heterogeneous evidence pieces and reason over their logical relationships; single‑evidence judgments are often insufficient, as illustrated by the provided examples.

The GEAR model (Graph‑based Evidence Aggregating and Reasoning) addresses this by constructing a fully connected evidence graph, applying graph‑attention mechanisms to aggregate evidence fragments, and using a classifier to label claims, while leveraging BERT embeddings for robust sentence representations.

Evaluations on the FEVER dataset show a 3‑4% accuracy gain over previous methods, especially on samples that require multi‑evidence reasoning.

Beyond fact‑checking, GEAR can be applied to detect fake information on online platforms, enhance knowledge‑graph completion, improve open‑domain QA, and be extended to verify user behavior for advertising purposes.

The authors also published a comprehensive survey on graph neural networks and will present detailed technical insights in an upcoming live session scheduled for June 23 at 19:00.