Knowledge-Based Misinformation Detection

The spread of misinformation on social media significantly impacts the democratic processes, disease control and economy. However, the amount of social media posts exhausted the fact-checking resources and brought a new major challenge to government responses worldwide. An artificial intelligence (AI) based automatic misinformation detection algorithm is required to minimise the human fact-checking efforts and enable large-scale misinformation detection and analysis. This project aims to develop a knowledge-based machine learning model (i.e. an algorithm that is trained to recognise patterns in the data) to detect misinformation on social media. Misinformation detection and debunking often require extensive domain knowledge and investigation. The traditional style-based classification method, which is based on the set of style features extracted from the text, can not provide valid evidence of its detection. Our proposed model will incorporate knowledge from professional fact-checkers in order to provide evidence/explanation to justify the model predictions. The knowledge includes professional debunking articles and verified information from trusted sources (e.g. government policy documents or fact-checking websites). In addition, this project also will explore the knowledge inference approach (i.e. acquiring new knowledge from existing facts) to extend the existing knowledge in order to detect unseen misinformation that the professional fact-checkers have not yet debunked. The approaches are based on the latest advances in the AI landscape. This translates into three objectives: Objective 1: Develop an automatic knowledge graph building algorithm to build such data structure from professional debunking articles. The knowledge graph is a well-established way to store and represent knowledge. The project will experiment with text semantic triple extraction methods to build a misinformation knowledge graph from the text. Objective 2: Create a novel misinformation detection model using a knowledge graph. The main activity for achieving this objective is to explore the method to incorporate the knowledge graph into the misinformation classification model. Objective 3: Build an explainable knowledge graph inference model to automatically predict the missing relations between the entities. The predictions will be used to detect unknown misinformation (using the outcome model of Object 2). Another potential avenue that will be explored is a multimodal approach to misinformation detection. This means analysing the multiple representations of the information contained within the misinformation posts, such as textual and visual. Many of said posts contain multiple modalities and that could be essential for machine learning models to make correct predictions. The detection results will be evaluated on the COVID-19 misinformation data created during the EU WeVerify project, using standard metrics such as F1 measure. A qualitative user-based evaluation will also be carried out with journalist and fact-checker users of the InVID-WeVerify verification plugin.