I have previously approached similar projects by using unsupervised learning using K-means clustering. That would be a pretty cool experiment and maybe the algorithm even comes up with a better categorization 
You can read about such an approach here https://towardsdatascience.com/applying-machine-learning-to-classify-an-unsupervised-text-document-e7bb6265f52
Just like @EricGT said: don’t be afraid to iterate, but close enough is close enough, and maybe have some TL3 users ready to re-categorize where necessary.