Exploring Language-appropriate Inductive Biases for Emotion Detection

Emotions are an integral element of human nature and are expressed in various situations in one’s life. These situations can stir a diverse range of feelings and emotions, e.g., from fear and sadness to joy and trust. When facing these circumstances, many people turn to online social platforms to express their emotions. Modeling emotions is extremely difficult and therefore emotion-related tasks are vital for assessing and improving the language understanding capabilities of large language models. This dissertation focuses on improving the understanding of emotions at scale. We propose novel algorithms and deep learning approaches to (1) detect expressions of emotions (although often not explicit) in a text and characterize their intensity and polarity; (2) identify the triggers (e.g., the topic, situation, or event) causing the emotions; and (3) predict how the emotions change over time and analyze emotion deviation (i.e., the varied emotions that people express towards the same trigger, e.g., two opposing signals – elation and sadness or rage – towards the same event).