Topic Modeling on Podcast Short-Text Metadata

Podcasts have emerged as a massively consumed online content, notably due to wider accessibility of production means and scaled distribution through large streaming platforms. Categorization systems information access technologies typically use topics the primary way organize or navigate podcast collections. However, annotating podcasts with is still quite problematic because assigned editorial genres are broad, heterogeneous misleading, data challenges (e.g. short metadata text, noisy transcripts). Here, we assess feasibility discover relevant from metadata, titles descriptions, using topic modeling techniques for text. We also propose new strategy leverage named entities (NEs), often present in Non-negative Matrix Factorization (NMF) framework. Our experiments on two existing datasets Spotify iTunes Deezer, dataset an service providing catalog podcasts, show that our proposed document representation, NEiCE, leads improved coherence over baselines. release code experimental reproducibility results ( https://github.com/deezer/podcast-topic-modeling ).