Topic Modeling and Progression of American Digital News Media During the Onset of the COVID-19 Pandemic

Xiangpeng Wan, Michael C. Lucic, Hakim Ghazzai, Yehia MassoudarXiv 2021 · arXiv

Introduction

The onset of the COVID-19 pandemic triggered a massive "infodemic"—a deluge of digital news that made it nearly impossible for individuals to maintain a holistic view of the crisis. This challenge was exacerbated by institutional biases, where different media outlets framed the same global event through varying ideological lenses. To address this, researchers developed an automated Natural Language Processing (NLP) pipeline designed to distill thousands of articles into manageable summaries and track the evolution of topic clusters over time.

Methodology

The proposed system employs a multi-stage architecture to transform raw news text into high-level trend analysis.

Summarization and Clustering

The pipeline evaluates two primary summarization techniques: the unsupervised TextRank algorithm, which utilizes SBERT embeddings to create semantic networks, and the supervised BART (Bidirectional and Auto-Regressive Transformers) model. BART, a denoising autoencoder, proved superior in capturing the core essence of news stories.

Following summarization, the system treats article summaries as nodes in a graph, with edges defined by the cosine similarity of their embeddings. Rather than using traditional centroid-based clustering like K-means, the researchers applied Louvain and Leiden community detection algorithms. This graph-theoretic approach better handles the dynamic, evolving nature of news threads.

Overview of the automated NLP pipeline for text summarization and topic modeling.

Topic Modeling

Once clusters (communities) are identified, the BART algorithm is utilized again to generate descriptive "community tags," effectively labeling the distinct news threads automatically.

Data and Performance

The study analyzed a custom-scraped corpus of 48,765 articles from ten major US media outlets (including The New York Times, CNN, and The Wall Street Journal) published between January 1 and April 2, 2020.

  • Summarization: BART achieved a ROUGE-L score of 47.09 on a benchmark NYT dataset, significantly outperforming extractive methods.
  • Clustering: The Louvain method reached a modularity score of 0.6381, indicating robust community structures. A similarity matrix confirmed that inter-cluster similarities remained low (typically below 0.6), validating the pipeline's ability to isolate distinct topics.

Temporal Topic Evolution

A key finding of the research is the dramatic shift in the media landscape around February 20, 2020. Prior to this date, coverage was dominated by international health concerns centered on Wuhan. Following the detection of community spread within the United States, the focus pivoted sharply toward domestic health, the US economy, and political ramifications.

Distribution of news categories from January to April 2020, showing the shift from international health to domestic issues.

By early April, domestic health issues accounted for over 40% of all articles. The study also noted "institutional fingerprints": while The New York Times and CNBC focused heavily on financial impacts, CNN and The Washington Post exhibited nearly identical topic distributions.

Conclusion and Future Work

While the pipeline successfully automates the distillation of news, some limitations remain. High-level categorization (e.g., "Economics Domestic") still requires manual tagging, a step the authors hope to automate with future classifiers. Additionally, moving from static datasets to real-time distributed systems is necessary for live news monitoring. Future research will also focus on quantifying sentiment and political "spin" to provide deeper insights into media bias.

Summary updated

BibTeX
@misc{wan2021topic,
  author = {Xiangpeng Wan and Michael C. Lucic and Hakim Ghazzai and Yehia Massoud},
  title = {{Topic Modeling and Progression of American Digital News Media During the Onset of the COVID-19 Pandemic}},
  year = {2021},
  eprint = {2106.09572},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.09572},
}
Figure