SocialVisTUM: An Interactive Visualization Toolkit for Correlated Neural Topic Models on Social Media Opinion Mining

Proceedings of Recent Advances in Natural Language Processing, pages 475–482 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_054 475 SocialVisTUM: An Interactive Visualization Toolkit for Correlated Neural Topic Models on Social Media Opinion Mining Gerhard Hagerer2, Martin Kirchhoff 3, Hannah Danner2, Robert Pesch3 Mainak Ghosh1, Archishman Roy2, Jiaxi Zhao2, Georg Groh2 1Max Planck Institute for Innovation and Competition 2Technical University of Munich 3inovex GmbH {gerhard.hagerer,hannah.danner,archishman.roy,jiaxi.zhao}@tum.de,grohg@mytum.de {martin.kirchhoff,robert.pesch}@inovex.de,mainak.ghosh@ip.mpg.de Abstract Recent research in opinion mining proposed word embedding-based topic modeling meth- ods that provide superior coherence compared to traditional topic modeling. In this paper, we demonstrate how these methods can be used to display correlated topic models on social me- dia texts using SocialVisTUM, our proposed interactive visualization toolkit. It displays a graph with topics as nodes and their correla- tions as edges. Further details are displayed in- teractively to support the exploration of large text collections, e.g., representative words and sentences of topics, topic and sentiment distri- butions, hierarchical topic clustering, and cus- tomizable, predefined topic labels. The toolkit optimizes automatically on custom data for op- timal coherence. We show a working instance of the toolkit on data crawled from English so- cial media discussions about organic food con- sumption. The visualization confirms findings of a qualitative consumer research study. So- cialVisTUM and its training procedures are ac- cessible online1. 1 Introduction Web sources, such as social networks, internet fo- rums, and customer reviews from online shops, pro- vide large amounts of unstructured text data. Along with the steady development of new platforms and the increasing number of internet users, the interest in methods that automatically extract the expressed opinions along with the corresponding topics and sentiments in text data has increased in recent years. Scholars and organizations from different fields can utilize such methods to identify patterns and gener- ate new insights. Examples are opinion researchers investigating current opinions on political and so- cietal issues, consumer researchers interested in 1https://github.com/ghagerer/ SocialVisTum Figure 1: SocialVisTUM applied to our use case or- ganic food - The topics, their occurrence (in brackets) and respective correlations. consumers’ beliefs about the consumption and pro- duction of goods (Danner et al., 2020), and market- ing managers curious about the public perception of their products and services (Berger et al., 2020; Murphy et al., 2014). (Kirchhoff, 2019) These domain-specific use cases are of interest for research disciplines which taken by itself are not directly related to natural language processing (NLP). Consequentially, there is a constant need to provide state-of-the-art NLP methods such that domain researchers from other fields can take ad- vantage of them. The requirements therefore are simple usage, automatic hyperparameter optimiza- tion, minimal effort for manual labeling of text data, and built-in visualizations to give an abstract overview of the discussed topics and their relation 476 with each other. While these practical requirements are impor- tant for domain experts, modern opinion mining approaches target specific machine learning objec- tives. Recently, there is a trend towards unsuper- vised neural methods for opinion target detection. Attention-based aspect extraction (ABAE) enables clustering of …

Content truncated to 500 words