I've done this professionally in a couple different settings, from building topic classifiers for news events (it is sometimes hard to know when one news event should stop and another start) to creating tagging systems for audio recordings of group conversations (where topics often merge in and out of each other, often within a single sentence).

I'm currently working on classifying non-speech, non-musical sound and it can be useful to piggyback on an existing knowledge system, though they tend to be industry-specific. As an example, Google's ontology for sound identification [1] is a nice starting point for general classification, whereas the taxonomy [2] used by the audio post-production industry (sound effects, foley, etc) is structurally quite different (which isn't surprising, but it sure is fun!). From a totally different field (electro-acoustic composition), the work of Michel Chion and Pierre Schaeffer [3] add psychoacoustic elements to more traditional measurable characteristics, i.e. how the sound is perceived and comprehended is just as important as its medium of travel and its source. It is helpful to see what others have done before you so you can pick and choose elements of their work to incorporate into your own.

1: https://github.com/audioset/ontology

2: https://docs.google.com/spreadsheets/d/1b2UhKpcOAE-jd1edOsxC...

3: [big pdf!] https://monoskop.org/images/0/01/Chion_Michel_Guide_To_Sound...