Projekt

History Podcast - Speech to Text to Temporal Network

Our project revolves around the history podcast „Geschichten aus der Geschichte“. This podcast is structured in a unique way, by regularly including references to past episodes. Because of that, the podcast as a whole has a graph-like structure. This is however hidden as there isn’t a formal documentation of these references. The general goal of the project is to make this hidden structure visible, both to uncover latent knowledge as well as providing an overview of the contents and relationships of the individual episodes. To achieve this goal, preexisting tools are combined with our own programming to customize the processes for the unique requirements of this specific use case. Foremost to mention is OpenAI’s Whisper, a speech-to-text conversion library, which was indispensable to create our data base to work from. With that, the explicit mentions could be extracted, and a foundational graph was created. This graph can be viewed and interacted with in Neo4j as the NoSQL database of our choice. The second pillar of the project is the Flair library in combination with the "ner-english-ontonotes-large" model, which was used to provide an extensive NER modelling with a total of 18 classes, which was applied to a total of 422 episodes. We started to take the “event” entity into account to increase the number of already captured entities from the episode descriptions, like persons or locations, and define implicit mentions based on episodic overlaps concerning these entities. There is a huge potential to build on the currently realized features, in particular the extension and improvement of extracted relationships in exchange with the podcast owners and their community.