Factual Error Correction of Claims from Wikipedia by Question Answering

Abstract

In this thesis, we identify a need for improved automated tools for fact-checking due to an increase in disinformation and misinformation. Furthermore, to create trust in the tools created we identify a need for these tools to incorporate some explainability in their determinations. As such we propose in this project "question answering based factual error correction" using a novel approach. Applying question generation on claims in tandem with question answering using the claims and provided evidence as context, we are able to both verify and correct claims. The generated corrections provide us with explainability and in conjunction with a verification model we can verify both the original and corrected claims. Our model is shown to be able to reduce refuted claims verified in the FEVER dev data set by 50% using this approach when compared to resulting verification using no corrections, while also having little effect on supported claims. This thesis also investigates what makes for high-quality questions in the fact-checking and correction domain using feature analysis and selection. We show that there exists no one feature that is the most important in making high-quality questions, but rather a combination of features, including downstream task features.