An Annotator Solves the Cold-Start Problem by Using Emotional Label Replacement and Consistent Selection Mechanism

Emotional annotators play an important role in the construction of emotional data sets. However, the construction of annotators often depends on the existing annotation data. When the annotation data does not exist in the required annotation category, the annotator will face a cold-start problem. In view of the characteristics that emotions can be expressed in discrete and continuous forms, we use the Pleasure, Arousal, and Dominance (PAD) emotion dimensional theory to propose a discrete emotional label replacement method to deal with the cold-start of the discrete emotional annotators. In order to compensate for the subtle differences between the replacement emotions, we propose a consistent selection mechanism to select the model that is finally used to automatically annotate the data sets to be annotated as an emotional annotator during the process of model training. Experiments performed on different aspects verify the effectiveness and universality of our label replacement and selection mechanism on solving the annotators' cold-start problem by comparing with some popular zero-shot methods.