In an era when we generate and disseminate information much faster than we can possibly consume, the need for factual verification has become more important than ever to help ensure the integrity of our information ecosystem. As automatic fact-checking systems emerge as an important research topic, the curation of reliable and trusted knowledge sources is equally essential. Over the past decade, organisations have curated high-quality KGs of various scales for diverse purposes. The wide curations of KGs make them a naturally good candidate as a reliable source for fact-checking and combating the spread of misinformation. However, because most KGs are built automatically or semi-automatically from existing texts or databases, they often suffer from incompleteness. Addressing this issue—known as knowledge graph completion—is essential for producing denser, higher-quality KGs that can enhance downstream applications.

This thesis investigates how KGs can support automated fact-checking by addressing two key challenges: improving KG completeness and leveraging KGs within fact-checking systems. First, this thesis recognises that KGs from different sources are often complementary and develops methods for knowledge graph fusion, integrating multiple KGs to enrich knowledge and reduce incompleteness. Building on these enhanced KGs, the thesis explores their application in automated fact-checking with minimal human intervention. An unsupervised KG-LLM pipeline that generates claim-evidence pairs simulating diverse fact-checking scenarios is proposed. This thesis then proposes to utilise the generated data from KG to 1) enhance the generalizability of existing fact-checking models and 2) enable a self-supervised pipeline where KGs serve as evidence sources for fact-checking tasks. Ultimately, the thesis aims to further automate and advance the curation of more complete, high-quality KGs and demonstrate their practical uses in improving factual verification systems without human involvement.