Using Knowledge Graphs to Detect Partisanship in Online Political Discourse

Existing methods for detecting partisanship and polarization on social media focus on either linguistic or network aspects of online communication, and tend to study a single platform. We explore the possibility of using knowledge graph embeddings to detect and analyze partisanship in online discourse. Knowledge graphs can potentially combine linguistic and network information across multiple platforms to enable more accurate discovery of a political dimension in online space. We train embeddings on heterogeneous graphs with different combinations of information text, network, single- and multi-platform information. Building on previous work, we develop a semi-supervised approach for uncovering a political dimension in the embedding space from a handful of labelled observations, and show that this method enables more accurate differentiation between liberal and conservative Twitter accounts. These results indicate that knowledge graphs can potentially be useful tools for analyzing online discourse.