"The Absence of Evidence is Not the Evidence of Absence": Fact Verification via Information Retrieval-Based In-Context Learning

Payel Santra, Madhusudan Ghosh, Debasis Ganguly, Partha Basuchowdhuri, Sudip Kumar Naskar

Abstract: Fact verification, the process of estimating the truth of a particular claim, plays an important role to automatically filter out misinformation and fake news in this age of information overload. Existing approaches of fact verification involve a supervised approach that relies on the existence of manually assessed resources in the form of claims and their truth or falsity. However, constructing such datasets requires significant manual effort, and, hence suffers from lack of scalability. Instead, in this study we demonstrate that an unsupervised non-parametric approach of using 0-shot or k-shot in-context learning turns out to yield results that, in terms of effectiveness measures, are close to that of a standard supervised method that involves fine-tuning a pre-trained transformer. In particular, we demonstrate that sentences extracted from webpages or Wikipedia, similar to the claim that is to be verified, turn out to be useful prompts towards effectively guiding a large language model-based decoder for this particular task.