I think the trick would be to suggest suitable pictures per paragraph based on NLP. If to take the highest rated content, there would already be a (biased) learning set available.