Checking Method for Fake News to Avoid the Twitter Effect

摘要

The recent blocking of President Trump's twitter account has raised awareness of the danger of the impact of fake news and the importance of detecting it. Indeed, if one can doubt information, ignoring what is true or false it can lead to a loss of confidence in the decisions and other dangers. The objective of this paper is to propose an automatic method for fact checking using Knowledge Graphs, such as Wikipedia. Knowledge Graphs (KGs) have applications in many tasks such as Question Answering, Search Engines and Fact Checking, but they suffer from being incomplete. Recent work has focused on answering this problem with an abstract embedding of the KG and a scoring function, yielding results that are not easily interpretable. On the other hand, Path Ranking methods answer this problem with deductions represented by alternative paths in the KG, easily understood by a human. Favoring the Path Ranking approach for its interpretability, we propose an attention-based Path Ranking model that uses label information in the KG, making the model easily transferable between datasets, allowing us to leverage pretraining and demonstrate competitive results on popular datasets.