Sounds good. How do you handle cases where the same document fits multiple categories (ie belongs in multiple locations of your hierarchy at once)?
Ps. Another question: can you tell which nlp libs are best for this purpose in your experience (and how do you eventually search the generated index?)