On End-to-end Automatic Fact-checking Systems

Abstract

The emergence of social media has aided the spread of nonfactual information across the internet, and organizations are combating disinformation by performing manual fact-checking. Due to the massive amount of online information, the automation of this process has recently gained great interest. Previous works have formulated several automatic fact-checking tasks, and explored machine learning and natural language processing approaches to the problems. In this thesis we follow this line of work, aim to build a fully-working automatic fact-checking system, and study methods for improving its fact-checking abilities. First, we introduce an end-to-end automatic fact-checking framework that integrates multiple previously studied subtasks to predict the factuality of given claims while providing supporting evidence. Next we explore the use of multi-task learning for improving factuality predictions. Finally, we devise methods for extracting temporal structure from news documents to aid the fact-checking process.