[Paper Review] Dynamic topic modeling of the COVID-19 Twitter narrative among U.S. governors and cabinet executives

Hao Sha, Mohammad Al Hasan+2 more|arXiv (Cornell University)|Apr 19, 2020

Misinformation and Its Impacts17 references48 citations

TL;DR

The paper applies a Hawkes binomial topic model to COVID-19 tweets from U.S. governors and cabinet members to track evolving sub-topics and infer influence networks.

ABSTRACT

A combination of federal and state-level decision making has shaped the response to COVID-19 in the United States. In this paper we analyze the Twitter narratives around this decision making by applying a dynamic topic model to COVID-19 related tweets by U.S. Governors and Presidential cabinet members. We use a network Hawkes binomial topic model to track evolving sub-topics around risk, testing and treatment. We also construct influence networks amongst government officials using Granger causality inferred from the network Hawkes process.

Motivation & Objective

• Analyze COVID-19 related Twitter narratives of U.S. governors, the President, and cabinet members from Jan 1 to Apr 7, 2020.
• Develop a Hawkes binomial topic model (HBTM) to capture evolving sub-topics around risk, testing, and vaccination/treatment.
• Construct influence networks via Granger causality inferred from the network Hawkes process.
• Identify phase-based shifts in topics and map cross-party and cross-state influence patterns.

Proposed method

• Use a Hawkes binomial topic model (HBTM) with intensity lambda_s(t,m) to model spontaneous and triggered tweets.
• Represent tweets as bags of words with a Binomial distribution for marks and a network Hawkes process for triggering between officials.
• Estimate branching probabilities q_ij to cluster tweets into dynamic topic groups over time.
• Restrict dictionary to the top W frequent words after stop-word removal; estimate spontaneous rates non-parametrically.
• Infer Granger-causality-based influence networks from triggering between officials and visualize cross-party interactions.

Experimental results

Research questions

• RQ1 What are the evolving sub-topics within COVID-19 tweets by U.S. governors and cabinet members from Jan 1 to Apr 7, 2020?
• RQ2 How do topics such as risk, testing, and vaccination/treatment emerge and change over time?
• RQ3 Which officials influence others in the Twitter narrative, and how are these influences distributed across party lines and states?
• RQ4 Can Granger causality inferred from the network Hawkes process reveal cross-government influence patterns during the pandemic?

Key findings

• Four temporal phases of narratives were identified, with a shift from low to high risk and rising emphasis on testing and live updates.
• Sub-topics around risk, treatment, and testing were detected, showing increasing attention to testing capacity and vaccine development over time.
• Influence networks show cross-party triggering and prominent roles for certain governors and officials, with the President having high spontaneous activity but limited cross-excitation influence.
• Phase-specific Granger networks reveal interconnected influence across parties and geographic clustering linked to state size and activity.