Technology to measure chance of news being fake
Posted: 26th June 2018
A ‘FakeNewsRank’ system that could show the likelihood of facts, articles, authors and websites being fake is to be developed with funding from the University of Oxford's EPSRC Impact Acceleration Account (IAA).
Fake news items on the web influence many people and have started to become a huge problem in important political and economic decisions, such as Brexit and the presidential elections in the USA. For this reason, there is major interest from governments and web companies alike (for example, social networks, web search companies and online newspapers) to detect and remove fake news. Clearly, this cannot be done manually anymore, given that there are more than a billion daily active users on Facebook. It is a natural step to think about developing artificial intelligence technologies to detect fake online news.
The one-year FakeNewsRank project led by Professor Thomas Lukasiewicz will be a step towards a spin-out company developing such technology. The goal of the project, which is about to start, is firstly to create a proof-of-concept demonstrator for computing a fake news score, denoted FakeNewsRank, for facts, articles, authors and websites. It will measure their likelihood of being fake, and identify any still existing key gaps or challenges. The project will also produce datasets (for two different domains, such as politics and the economy) of sample blog messages and news articles, containing true and fake facts, along with their authors and web addresses, and background knowledge graphs underlying these datasets.
The demonstrator will be based on Thomas and his group’s deep-learning-based system for extracting facts from plain text, which extracts facts relative to a background knowledge graph from plain natural language text on a web page. Furthermore, the project will also be based on the group’s deep-learning-based system for ontology reasoning, which will be used to determine the likelihood of truth of an extracted fact relative to the underlying background knowledge graph. The FakeNewsRank computed by the demonstrator will be based on this likelihood of extracted facts and on other parameters.