Verif-AI

How can we scale AI-based misinformation detection to work multilingually, over the noise of social media?

The web is rife with misinformation, which affects the lives of those who have come to use it as an authoritative source of information. The veracity of information spreading through social media can sometimes be hard to establish and the deliberate or accidental spread of false information, especially during natural disasters or emergencies, is quite common. Malicious actors are now also well-versed in using the web to shape and manipulate public opinion, and thus democracy, in many countries, including Denmark. To detect false claims automatically, large volumes of user-generated content need to be analysed quickly, yet complex real-time analytics are a major outstanding challenge.

The project proposes to grow a set of known text-based and network-based Artificial Intelligence (AI) techniques for fake news detection and fact verification from beyond their locales in the English-speaking world, to a broad multi-lingual, multi-social approach.

The routes for investigation comprise:

• Fact extraction and verification: identifying the claims in text by extracting them into a computational representation, and then comparing them with knowledge bases, to determine what may be verified, what may not, and what may be refuted. This builds upon core natural language processing (NLP) techniques, adapting them to both the noise in social media (a challenging but prevailing environment) and also to new languages (including Danish)
• Stance detection: veracity can be approximated from the way others react and orient to new claims. These reactions are observable in the form of social media and web comment replies. This is helpful for tackling emerging claims that aren‘t in knowledge bases yet (e.g. ‘Rådhuset brander!‘). It requires rapid detection of claims; accurate detection of the attitude – or stance – that a commenter takes; and a model for interpreting the flow of reactions and stances made. While preliminary tools for this exist, they are not multi-lingual, and only operate on a small range of stories. This project will broaden the scope of stance detection to be cross-lingual, and ground it to veracity detection.

This will lead to a more robust understanding of the influence and propagation of fake news on social media, to better and to new tools for working with this phenomenon, while using Denmark as one of the non-English test cases, bringing these advanced tools to our country.