It is because the level of feasible phrase sequences boosts, plus the designs that tell effects come to be weaker. By weighting words in a nonlinear, distributed way, this model can "study" to approximate words and not be misled by any unidentified values. Its "comprehension"