Establishing Relationships between Emotion Taxonomies Using the Vector Space Model

Due to different aspects that emotion-oriented research looks to capture, the emotion taxonomy used often differs among research efforts. Therefore, it is hard to coordinate the research efforts using different emotion taxonomies. On the other hand, due to the multiplicity of “emotion”, emotion annotations more naturally fit the paradigm of multi-label classification since one instance (such as a sentence) may evoke a combination of multiple emotions. We thus propose bridging the gap between emotion taxonomies in the multi-label domain by leveraging the Vector Space Model and crowdsourcing. The relationships between source emotion taxonomy and target emotion taxonomy are formalized as a transformation mapping, which is established using the gold emotion annotations in the source taxonomy and the crowdsourced emotion annotations in the target taxonomy. Using the established mapping, associated emotions in the target taxonomy for an instance can be directly obtained according to its associated emotions in the source taxonomy. Experimental results on the real-world data demonstrate that the mapping established using the proposed models enables the gold emotions in the target taxonomy to be effectively estimated.