Twitter
Wang et al. [12] noted the lack of manually annotated
training data available for emotion recognition and proposed
Twitter as a solution. They considered the problem of categorising
tweets into 7 discrete emotional categories, and
noted their intention of applying the work to other domains
such as blog posts. Using a large database of 2.4 million
tweets, they demonstrated an impressive 65.57% accuracy
in a 7-emotion category task. Categorical emotions were assigned
using the presence of one of 131 emotional hashtags
that appeared at the end of each tweet such as `#nervous'
or `#love'. Hashtags were a convincing approach for discrete
emotional labels, as hashtags are themselves a user's
categorisation of their own tweet. Pak et al. [8] investigated
the problem of classifying the sentiment of twitter posts us-