PNNL

Abstract

The ability to detect and characterize online information campaigns will rely on the ability to infer shared underlying narrative viewpoints of online content. It is challenging to discover and categorize groups of documents which share both an underlying topic area and corresponding viewpoints as existing methods typically can cluster documents by topic but struggle to differentiate between different points of view. In the context of social media, this task becomes more challenging due to the inherent characteristics of social media texts, which are noisy, short, and often provide very little context. Yet, due to the widespread prevalence and traction of harmful misinformation in online media, the development of narrative viewpoint detection approaches is crucial to eventually enable the identification of information campaigns, the characterization of their sources, and their evolution behavior in online platforms. Our work proposes a weakly supervised contrastive representation learning approach to infuse latent text representations with viewpoint information. Our approach focuses on proxy signals observed through social interaction networks to encode viewpoint similarity in the semantic latent space. Using two Twitter datasets related to COVID-19 discussions, we demonstrate the ability of our models to successfully separate tweets by their narrative viewpoint in various topics compared to baseline pre-trained embeddings.