Exposing a Set of Fine-Grained Emotion Categories from Tweets

An important starting point in analyzing emotions on Twitter is the identification of a set of suitable emotion classes representative of the range of emotions expressed on Twitter. This paper first presents a set of 48 emotion categories discovered inductively from 5,553 annotated tweets through a small-scale content analysis by trained or expert annotators. We then refine the emotion categories to a set of 28 and test how representative they are on a larger set of 10,000 tweets through crowdsourcing. We describe the two-phase methodology used to expose and refine the set of fine-grained emotion categories from tweets, compare the inter-annotator agreement between annotations generated by expert and novice annotators (crowdsourcing) and show that it is feasible to perform fine-grained emotion classification using gold standard data generated from these two phases. Our main goal is to offer a more representative and finer-grained framework of emotions expressed in microblog text, thus allowing study of emotions that are currently underexplored in sentiment analysis.