Using Co-word Network Community Detection and LDA Topic Modeling to Extract Topics in TED Talks

Two topic detection techniques-co-word network analysis and topic modeling-were applied to extract topics in the Ted Talks. Ted Talks was chosen for its enormous impact worldwide and the rich descriptive data accompanying each talk that allow us to compare the topics resulting from different methods. The co-word network was built based on the "related_tags" field so that modularity analysis can be performed to classify the tags according to their co-occurrence patterns. Topic modeling was applied to the description field and the full-text transcript separately to detect the topics present in the free-text. The results of network modularity analysis revealed 13 interpretable topics consisting of closely knitted tags. Topic modeling generated 25 topics for the description and 40 for the transcript, respectively. Our results showed that both topic extraction methods were able to successfully identify the range of topics in the TED Talks. While the co-word network gave a broad overview and afforded visualization, the topic model revealed topics with greater granularity. We compared the semantics of the topics produced by different methods and discussed the methodological implications of our research.