Computational Fact Validation from Knowledge Graph using Structured and Unstructured Information

In today's world, data or information is increasing at an exponential rate, and so is the fake news. Traditional fact-checking methods like fake news detection by experts, analysts, or some organizations do not match with the volume of information available. This is where the problem of computational fact-checking or validation becomes relevant. Given a Knowledge Graph, a knowledge corpus, and a fact (triple statement), the goal of fact-checking is to decide whether the fact or knowledge is correct or not. Existing approaches extensively used several structural features of the input Knowledge Graph to address the mentioned problem. In this work, our primary focus would be to leverage the unstructured information along with the structured ones. Our approach considers finding evidence from Wikipedia and structured information from Wikidata, which helps in determining the validity of the input facts. As features from the structured domain, we have used TransE embedding considering components of the input fact. The similarity of input fact with elements of relevant Wikipedia pages has been used as unstructured features. The experiments with a dataset consisting of nine relations of Wikidata has established the advantage of combining unstructured features with structured features for the given task.