Toward Automatic Misinformation Detection Utilizing Fact-checked Information

Abstract

We proposed a new task FCCKB: Fact-checking by Claim Knowledge Base. The goal was to fact-check a sentence utilizing verified claims stored in the database. To retrieve relevant claims from the large database, we proposed applying Semantic Role Labeling(SRL) on the input sentence having rich semantics and then encoding the results to get fine-grained sentence embeddings. That improved semantic matching between the input sentence and the relevant claims. We used three sentence encoders for sentence encoding. In FEVER dataset, precision and recall was improved by more than 5 percent after SRL was applied.

Tasks

Reproductions