ing an automatically tagged corpus of tweets. Instead of
hashtags, the presence of emotions was used to tag tweets
as having either `positive' or `negative' sentiment, with `neutral'
training data sourced from newspaper twitter accounts.
Their system achieved around 60% accuracy for the threeclass
problem using a Nave Bayes classier trained on POS
tags and n-grams. As in [12], accuracy was found to increase
monotonically with the training dataset size, albeit with diminishing
returns. Uni-grams and bi-grams used together
were found to be best, a nding that agreed with