From Single-Task to Multi-Task: Unveiling the Dynamics of Knowledge Transfers in Disinformation Detection

Abstract

The widespread dissemination of misinformation and fake news across digital platforms poses significant societal challenges. Modern detection approaches employ multi-task learning to harness the relationships between disinformation-related tasks (e.g., stance detection and rumor classification), aiming to enhance overall detection performance. However, knowledge transfer between tasks can lead to performance degradation (negative transfer) rather than improvement (positive transfer). Despite designing efforts towards more complex architectures, the underlying mechanisms and reasons for this phenomenon remain unclear. In this paper, we directly target this problem by examining similarities and differences between models trained under single-task and multi-task settings. Specifically, we consider a comprehensive set of disinformation-related tasks, including Sentiment Analysis, Fake News Detection, Stance Detection, and Topic Detection, and pioneer the utilisation of explanations to uncover the differences between models trained under single-task and multi-task settings. We find that positive transfer occurs across several combinations of the examined tasks. In particular, our findings reveal that positive transfer refines the knowledge that can already be learnt in single-task settings by incorporating additional patterns from other tasks. Conversely, negative transfer significantly undermines models' knowledge to the extent that their explanations are equivalent to a random perturbation of the explanations generated by their single-task counterparts.

Keywords: Disinformation mining, Knowledge transfer, Explainable AI