GPT-Summarized Clusters in CiteSpace

A critical step in the process of reviewing and making sense the thematic landscape of a field of research is to synthesize a variety of research questions, findings, and uncertainties. This step is often cognitively demanding. Ideally, it would be nice to have multiple areas of expertise readily accessible. CiteSpace has so far facilitated the task by underlining the critical works that researchers should take into account according to structural and temporal properties such as works receiving rapidly increased attention and works with great boundary-spanning potentials. Here I'd like to share some examples of how GPT models can enhance this critical step.

The Summary Report function in CiteSpace automatically generates a report that highlights major clusters in terms of their properties and leading articles, including those with strong citation bursts and high betweenness centrality. We have seen how instrumental GPT models can be in generating meaningful cluster labels from citing titles and reference titles associated with a cluster. Now we take a small step further and capitalize on GPT to give us a cohesive summary of a cluster. The GPT-enhanced Summary Report function synthesizes prominent themes in a cluster. The example below illustrates the upgraded function. The report starts with an overview of the visualized network to provide a context for subsequent exploration and sense making tasks.

Here is the overview of a dataset used in the Duality of Citation blog.

A nice byproduct is we can easily switch to different languages. Here is the same visualization in Chinese.

As a user, you can select how many clusters you would like to be included in the report. Each cluster is summarized with the top 5 most influential citing articles, i.e., they cited the most number of references in the cluster. You can access them in more detail with their hyperlinks. The most interesting addition is the Summary section, which is entirely generated through the GPT API from CiteSpace.

Here is the same result presented in Chinese.

The same summary in English and Chinese side by side: