Robust n-Gram Model of Japanese Character and Its Application to Document Recognition