Handling Ambiguity in Emotion: From Out-of-Domain Detection to Distribution Estimation

Anonymous

TL;DR: Estimate emotion distribution with a reliable uncertainty measure.

Abstract: The perception and interpretation of speech emotion are highly subjective, resulting in inconsistent labels from human annotators. Typically, only data with majority-agreed labels are used to train emotion classifiers, which results in the exclusion of data without majority-agreed labels and poses challenges to the model's generalisation ability when ambiguous emotional expressions are encountered in test. To handle ambiguous emotional speech, three methods are studied in this paper. First, an approach based on evidence theory is introduced to quantify the uncertainty in emotion class prediction and detect utterances with ambiguous emotions as out-of-domain samples using the uncertainty score. Second, to obtain fine-grained distinctions among ambiguous emotions, we propose re-framing emotion classification as a distribution estimation task, where every individual label is taken into account in training, not just the majority opinion. Finally, we extend the evidential uncertainty measure for classification to quantify the uncertainty in emotion distribution estimation. Experimental results on the IEMOCAP and CREMA-D datasets show that our method produces effective emotion representations with a reliable uncertainty measure.

Paper Type: long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Languages Studied: English