Topic modeling looks at patterns of keywords that frequently occur together in a set of documents. They trained and compared topic models using "Latent Dirichlet Allocation," "Correlated Topic Models," "Hierarchical Dirichlet Process," and a range of other topics between five and fifty. 

They found similar results in topic contents and Wikipedia coverage of these topics across models and multiple runs. 

COVID-19 research in Wikipedia methodology: topic modeling clustering

SPECTER and k-means was used to cluster the papers. SPECTER is a new method to generate embeddings of documents based on its network of citations and a transformer language model. Every paper was embedded and then clustered with k-means, using k = 20. 

COVID-19 research in Wikipedia methodology: SPECTER

For text-based clustering, two methods were used: topic modeling and k-means using SPECTER embedding, which is a new method to generate document-level embedding of scientific documents. 

University of California, Berkeley

Researchers used clutering analysis using publication text and citation data. For publication text, they used text-based clustering. For citation data, they did citation-network clustering. 

COVID-19 research in Wikipedia methodology: clusters

Colavizza, G. (2020). COVID-19 research in Wikipedia. Quantitative Science Studies, 1-32. https://www.mitpressjournals.org/doi/full/10.1162/qss_a_00080 

COVID-19 research in Wikipedia

COVID-19 research in Wikipedia methodology: text-based clustering

For citation network clustering, researchers created a bibliographic couplic citation network using publications and references from Dimensions. Edges were then weighted through fractional counting and then clustered through the Leiden algorithm. Through this, they found that the largest 43 clusters accounted for half the nodes in their network. The largest cluster had 15,749 nodes. 

Learn Before

Related

Learn After