Computational Sarcasm Detection Examined Through the Lens of Psycholinguistics

Abstract

This thesis introduces a new framework that examines sarcasm detection from the combined angles of psycholinguistics and computational linguistics. It provides empirical evidence about various factors that cause sarcasm and affect the understanding of it in human communication. The thesis also uses those findings to deepen our understanding of computational sarcasm detection models. Our first contribution is to reveal how certain important factors relate to the production and interpretation of sarcasm. Two of our experiments examine sarcasm's relation to context, emotion, and communicative intent. We find that emotion and communicative intent are highly associated with the choice to use sarcasm. Using two further experiments, we clarify the relation between speaker perspectives and observer perspectives in navigating communicative situations involving sarcasm---a topic that is an important but neglected aspect of computational sarcasm detection. Our second contribution is to use the key factors just mentioned as a means of investigating the nature of the automatic sarcasm detection process. We report several patterns identified through three computational experiments, whose united goal is to reveal what kinds of information are encoded in sarcasm detection models, and whether they are similar to the information shown in human behavior. The first finding is that generalizable automatic sarcasm detection is difficult, due to the varied characteristics that sarcasm has. Additionally, we find that the incongruity between the speaker's affect and the content of the utterance makes the identification of sarcasm challenging both for human observers, and language models. Finally, we show that different amounts of context play different roles in computational sarcasm detection, and that this depends on the level of disagreement about sarcasm among human observers. Our last contribution in this thesis is to introduce a new framework that smoothly integrates experimental and computational investigation of sarcasm, which can be applied to other linguistic phenomena in future research. Our proposed framework, which is proven effective through multiple studies reported in this thesis, should bridge human communication and natural language processing, and contribute to a more multifaceted investigation of natural language in future work.