Sanja Sarić
Centre for Informationmodelling, University of Graz
Summary
The work of Sanja Sarić focuses on using networks to visualize topic modeling results in order to show the closeness between entities and the inclusion of further annotated data from the corpus. The network graphs were created in Gephi, and the nodes were colored based on calculated modularity. She is convinced that the network visualizations were insightful and compelling because they were able to demonstrate otherwise invisible relations. However, it is noted that the complexity of the visualizations presents a challenge for researchers unfamiliar with the method. The aim of the presentation was to follow-up with a discussion about the use of networks in combination with topic modeling and ways to improve their comprehensibility in the research community.
Keywords
topic modelling, multilingual data, network visualisation
Slides
Feedback
Linda FREYBERG stresses the importance of being unambiguous and transparent regarding the application’s terminology, e.g., the meaning of probability calculations or modularity groups in networks. In addition, even though topic labeling can be done fully automatically, it is advised to rely less on automation and to contemplate the use of domain-specific thesauri for this task.
Keith ANDREWS recommends contrasting the research with other tools that visualize networks or belong to the domain of Visual Document Analysis. He stresses that making the results as deterministic as possible is key, such as by applying t-SNE (t-Distributed Stochastic Neighbor Embedding) in addition to Gephi’s layout algorithms Fruchterman-Reingold and ForceAtlas.
Florian WINDHAGER proposes using word clouds on top of clustering methods, such as by integrating legends or mouse-over effects. Moreover, additional layers to information space seem sustainable; joining multiple and expressing them simultaneously color-wise may enhance comprehension.
Tiziana ALOCCI advises using a more up-to-date visualization software and also securing a more concrete access point into the visual itself, regardless of whether the implementation is primarily aimed at Romance studies research experts.
Resources
- PNNL: IN-SPIRETM - About IN-SPIRETM. https://in-spire.pnnl.gov/about.stm.