Abstract

An approach for determining emotional ambiguity in text data is described in this paper. The prediction confidences output from a text classifier are used to measure amount of ambiguity found in target entries. This measure can be used as a filtering mechanism to identify entries that require human feedback. This feedback loop can be implemented in a workflow which retrains a classifier model including newly disambiguated entries and resulting in a boost to classifier accuracy. This emotion ambiguity measure can be utilized to discover concrete emotional content in text data as well as reveal topics which do not have a concrete emotional consensus. 

Keywords

emotion, ambiguity, deep learning, natural language processing