This is due to the level of doable term sequences boosts, plus the designs that notify benefits turn into weaker. By weighting terms in the nonlinear, distributed way, this model can "discover" to approximate words and not be misled by any not known values. Its "understanding"