Generating explainable answers to fact verification questions

Online platforms such as social media and news sites have recently become vehicles for widespread misinformation, making it challenging for users to distinguish facts from hearsay. This can be addressed by the task of fact verification, which is aimed at automatically assessing the truthfulness of claims. Traditionally, fact verification has been cast as a binary classification task, whereby a given claim is labelled as true or false by a classifier (Thorne and Vlachos, 2018). More recently, it has also been formulated as a question answering task (Zhang et al., 2020, Jobanputra, 2019), thus providing end-users with the ability to pose their doubts as questions (e.g., "Does Ibuprofen make COVID-19 worse?") and retrieve answers in real-time. These answers can come in a short form (e.g., "No") or in a more informative form (e.g., "There is no evidence linking worse COVID-19 symptoms to Ibuprofen") that can include not only evidence from multiple sources but also explanations.

In this project, the PhD candidate will develop machine learning-based methods for fact verification, casting it as a question answering task. Along the way, the candidate will investigate neural architectures for machine reading comprehension (Liu et al., 2019, Qiu et al., 2019) as well as methods for information retrieval (for selecting relevant and reliable information sources) and natural language generation (for producing long-form answers to questions). Importantly, the candidate will propose approaches for making any generated answers explainable in the form of human-readable natural language text, enabling end-users to better understand and interpret the answers generated by the fact verification model (Kotonya and Toni, 2020).