Dynamic Topic Modeling of Video and Audio Contributions

Publikation: KonferenzbeitragPapierBegutachtung


This paper shows how topics and their temporal evolution in audio and video
broadcasts can be analyzed and visualized automated. For this purpose, Deep Learning systems such as "OpenAI Whisper" and "GPT3" are used to transcribe the audio data and extract the essential content per broadcast. The "BERTopic" method (Grootendorst, M. 2022) is used for dynamic topic modeling. The result is clusters of content ("topics") that are described and visualized using scatter plots, word clouds, and line charts. The method solves problems of topic modeling and enables the automated analysis of large amounts of data. A software prototype was developed that combines the sub-models and enables the analysis. The method is demonstrated using the example of Austrian TV
channel ServusTV's weekly commentary "Der Wegscheider" over a period of more than four years (2018-2022). It is shown that migration, "mainstream media," and the Covid19 pandemic are dominant topics. The time trend analysis illustrates how the COVID-19 pandemic increasingly crowded out the other topics from mid-2019. This method demonstrates how AI can be applied to journalistic work to enable the analysis and visualization of large data sets.
OriginalspracheDeutsch (Österreich)
PublikationsstatusVeröffentlicht - Dez. 2023
VeranstaltungSTS Conference Graz 2023; Critical Issues in Science, Technology and Society Studies - TU Graz, Graz, Österreich
Dauer: 8 Mai 202310 Mai 2023


KonferenzSTS Conference Graz 2023; Critical Issues in Science, Technology and Society Studies