Mostly #2 of what the other response says. The entire input to the NN is what it's learning off - the weights are static, you are changing the probabilities based on which context is provided and it's length.