Emotional classification in COVID-19: Analyzing Chinese microblogs with domain-adapted contrastive learning

Abstract

Emotion analysis for COVID-19 is a domain-specific task, such as the epidemic, which plays a significant part in scientific research institutions and governments to track the emotional changes and trends of society. When introducing general domain textual information, currently used techniques just concentrate on learning the domain-invariant information to reduce domain discrepancy but ignore the maximum use of domain-general information to solve the problem of domain-specific data scarcity. As a result of this inspiration, we develop a domain-adapted contrastive learning-based emotion classification model, which consists of three modules: text representation, emotion identification, and domain adaptation. In this model, the text representation module is used to obtain a representation of sentences, and then the domain adaptation module is employed to pull the representation space of domain-specific data and domain-general data to overcome domain discrepancy and ultimately achieve better performance in the emotion identification module. To fortify our model, we propose two different contrastive learning strategies in the domain adaptation module. Experimental results on the SMP2020-EWECT show that our two strategies achieve F-values of 66.28% and 67.39% respectively, which significantly outperform the baselines despite the scarcity of domain-specific data. Interpretability analysis further demonstrates that the model employing domain-adapted contrastive learning can better understand domain text emotions.