This paper describes a method to expose a set of categories that are representative of the emotions expressed on Twitter inductively from data. The method can be used to expand the range of emotions that automatic classifiers can detect through the identification of fine-grained emotion categories human annotators are capable of detecting in tweets. The inter-annotator reliability statistics for 18 annotators using different granularity of the emotion classification schemes are compared. An initial set of emotion categories representative of the range of emotions expressed in tweets is derived. Using this method, researchers can make more informed decisions regarding the level of granularity and representativeness of emotion labels that automatic emotion classifiers should be able to detect in text.
展开▼