Multi-Task Learning Framework for Stance Detection and Veracity Prediction

Tracking #: 2642-3856

This paper is currently under review
Fatima T. Alkhawaldeh
Tommy Yuan
Dimitar Kazakov1

Responsible editor: 
Guest Editors ST 4 Data and Algorithmic Governance 2020

Submission type: 
Full Paper
Detecting unverified information automatically has become an essential topic of research as many people get their news through online sources. Stance section and rumour verification tasks have gained increasing interest in recent research. Most of the existing models train both tasks separately or consider the sources contribute equally. This paper proposes a multi-task learning framework for jointly predicting evidence source, stance and veracity claim on news, to enhance the performance of veracity prediction. One of the main goals of this model is to generate the best conclusion from the available evidence in case of a lengthy article. Another goal is to detect each evidence stane toward a particular claim then qualify its confidence among conflict facts using a unified model that could be adapted to the news domain. This work implements an argumentation-based truth discovery approach to reason about contradiction beliefs, given by various sources with different reliability levels. Experiments on Emergent and SemEval 2019 Task 7 datasets show that this method outperforms previous methods on both stance classification and veracity prediction.
Full PDF Version: 
Under Review