It's not that the tag "excited" is preferred, but a probability that the message should in fact be classified as "excited" and not "annoyed."
Suppose I have 2 classifications for sentiment: "bullish" and "bearish." I then train a model in the Prediction API with even amounts of "bullish" and "bearish" training data. When I submit a message to Prediction API to get the sentiment, it reads the text and assigns a probability both a "bullish" and a "bearish" probability based on the words in the message. The sum of the probabilities will add up to 1.
So again, it's not that one label is preferred to another, but the probability of the message being "excited" is 3 times greater than it being "annoyed."