Obviously I'm not a $100M AI genius but haven't they tried transforming both training data and context into some normal form (whitespace neutral)?