"spam and impersonation is increasing on nostr recently… " - not in my hizzy. I trained against 4000 nostr posts to suppress bitcoin price articles, zapathons, spanish poetry, vulgar language and whetever else. https://cafe-society.news/nostrposts. Not good yet, but much better than IRL. Here is the winknlp compatible model if you're curious. https://github.com/colealbon/squelch/blob/50772762d7ed6b7c4f920903b243083972e069c2/src/defaultClassifiers.ts#L38