You'll get 80% of the benefit just by looking at word frequency, highlighting outliers and then a weight based on factors such as length and secret-sauce weighting.

Bonus points if you're using multiple categorizations (using different weights for different industries).

NLP / statistical stuff is fun ;)

Are you scanning / OCRing the documents? I never managed to get the OCR to be good enough for invoicing, there always had to be a manual process to fix the (machine-learning-flagged) errors.

Or don't you need accurate-to-the-cent invoices?