What is shown?

The visualization above shows every publication embedded in a 3D space where the position is determined by the topic of the publication. This means that papers that investigate the same research area are closer together than others.

How to navigate?

For desktop users it is possible to navigate with the mouse and pan with the arrow keys. The pan and zoom speed can be adjusted with the controls on the side to allow for more fine grained control. Users with a touch screen can navigate through the research as well. A user can switch to full screen mode (top right corner) if desired. By clicking on a category the category is highlighted in the publication cloud. When the user clicks on a paper, it and its five closest neighbors are highlighted and displayed in a list below. The papers are ordered by the distance from the selected paper which means that the more closely related paper will be above papers that investigate a different topic. It is also possible to click on a topic (list below) to highlight a specific research area.

What is the purpose?

The visualization of the current research provides useful insights on emerging topics. Users can easily navigate through the publications and find correlations between research fields and papers that they did not know before. The topic assignment of the papers are supporting researches to find relevant articles for their specific interests.

How are the embeddings generated?

We trained a neural network (a BERT transformer model) given a title and an abstract from a publication to predict if they are from the same publication or not. By randomly exchanging some title and abstracts during training, the network has to learn what the papers are about to be good at the prediction. It can then generate a 512 dimensional encoding of every publication. To visualize these generated embeddings, we reduced the dimensionality of each vector with T-SNE to a three dimensional space.

How are the topics generated?

To generate the different topics, we clustered the embeddings and then assigned a name to the topic that in our opinion best reflects the contents of the contained papers.