Modeling Emotion across Languages, Label Formats, and Linguistic Levels

Language-based emotion analysis finds itself in a paradoxical situation. In the past decades, a plethora of datasets have been created, covering diverse aspects of natural language and affective states. However, the considerable volume of resulting gold data is scattered across many design decisions in dataset creation, acting as sources of heterogeneity (e.g., different natural languages, linguistic units of different sizes such as words, sentences, and texts, the choice for a particular set of emotion variables, and the selection of a particular viewpoint such as the reader or the writer perspective on emotion). On the one hand, this proliferating heterogeneity makes it difficult to re-use existing datasets and software tools or to compare experimental results following different design decisions. On the other hand, this heterogeneity is empirically adequate and valuable from an application perspective. Thus, the solution to the heterogeneity problem cannot be to simply reduce the number of possible design choices through community-wide consensus. Rather, what is needed is a delicate balance between fostering the diversity of emotion data and developing new methods to tackle the resulting comparability issues. This interplay between diversity and comparability of emotion data is the focus area of this thesis and the seven studies compiled within it. The larger vision behind this dissertation is to arrive at a research landscape where diversity and comparability no longer act as antagonists and instead every new sample of annotated data, regardless of the specifics of its annotation design, benefits the endeavor of emotion analysis as a whole. While NLP is still far from achieving this goal, the presented research results, culminating in the establishment of an “emotion interlingua” in the final study, constitute a significant step in this direction.