Context-Aware Message-Level Rumour Detection with Weak Supervision

Social media has become the main source of all sorts of information beyond a communication medium. Its intrinsic nature can allow a continuous and massive flow of misinformation to make a severe impact worldwide. In particular, rumours emerge unexpectedly and spread quickly. It is challenging to track down their origins and stop their propagation. One of the most ideal solutions to this is to identify rumour-mongering messages as early as possible, which is commonly referred to as "Early Rumour Detection (ERD)". This dissertation focuses on researching ERD on social media by exploiting weak supervision and contextual information. Weak supervision is a branch of ML where noisy and less precise sources (e.g. data patterns) are leveraged to learn limited high-quality labelled data (Ratner et al., 2017). This is intended to reduce the cost and increase the efficiency of the hand-labelling of large-scale data. This thesis aims to study whether identifying rumours before they go viral is possible and develop an architecture for ERD at individual post level. To this end, it first explores major bottlenecks of current ERD. It also uncovers a research gap between system design and its applications in the real world, which have received less attention from the research community of ERD. One bottleneck is limited labelled data. Weakly supervised methods to augment limited labelled training data for ERD are introduced. The other bottleneck is enormous amounts of noisy data. A framework unifying burst detection based on temporal signals and burst summarisation is investigated to identify potential rumours (i.e. input to rumour detection models) by filtering out uninformative messages. Finally, a novel method which jointly learns rumour sources and their contexts (i.e. conversational threads) for ERD is proposed. An extensive evaluation setting for ERD systems is also introduced.