Add more word in model but not in dict

[wengxt]

based on the algorithm, it is only meaningful when a word exists in model.

It seems that we missed many words that actually present in the model so in some case we gives inaccurate prediction.