Knowledge Graph Fact Checking using Deep Learning

Abstract

In recent years, the use of knowledge graphs has become an essential part of information retrieval. These knowledge graphs are made up of billions of facts that are encoded in the form of RDF triples. They are automatically generated so that means they are prone to errors. Therefore, providing evidence for the validity of these facts can be of great importance. Such evidence can be in the form of text snippets that are extracted from reliable sources. In this thesis, we propose an approach that takes a fact in the form of an RDF triple and checks if it is true or not and provides textual evidence for the given fact. Our proposed approach works into two stages. First, for a given knowledge graph fact, it retrieves the top-k most relevant sentences to the fact from a background corpus. Second, using these retrieved sentences along with the input fact, it automatically verifies whether a fact is true or false by making use of newly developed deep learning models. We evaluate our approach and compare it to various baselines using FactBench, a publicly available benchmark for fact checking. Our experimental results demonstrate the merits of our proposed approach in validating knowledge graph facts.