Assign labels/create training data for a large dataset

[EY4L]

What is the most efficient way to process data to create training data if you have a very large dataset and classes, for example 800k rows, with 1.6k classes, training a text classifier?

Currently my approach is to create a category dictionary with all classes {'cat1': 0, ... , 'cat2':1} for the respective class being assigned, then doc=nlp(text) the example and assign the new category dictionary in the doc.cats dict.

I know there are the following speed improvements I can try, but are there any other approaches?
using nlp.pipe and passing a batch of texts

doc.cats dict only needs the class with a '1' value:
Current method:
training_data = [
("This is a sports article.", {"cats": {"sports": 1, "politics": 0, "technology": 0}}),
("This is a political article.", {"cats": {"sports": 0, "politics": 1, "technology": 0}}),
]

Faster approach
training_data = [
("This is a sports article.", {"cats": {"sports": 1}}),
("This is a political article.", {"cats": {"politics": 1}}),
]