Using Spacy for OCR data set with 'junk' tokens

Question

I am using Spacy Named Entity recognition to extract specific names from document image OCR. My training data sets comprise of up to 6000 documents, up to 4 pages each, annotati