PoliGuilt: Two level guilt detection from social media texts

Abstract

Guilt, a multifaceted emotion stemming from the realization of causing harm, intertwines with various aspects of human psychology and social interaction. This paper delves into the nature of guilt by developing an annotated dataset of 3,304 posts. Guilt detection is approached as a two-level classification task: first, distinguishing between guilt and non-guilt, and then categorizing guilt into the types “Anticipatory”, “Reactive”, and “Existential” based on psychological frameworks. Exploratory analyses are conducted to examine the contributions of post titles, self-text, and their combination as inputs to guilt detection algorithms. Various learning approaches were employed, including traditional machine learning, deep learning models, and transformers, to ensure quality and efficacy. The findings indicate that while simple methods using only unigrams can distinguish between texts expressing guilt and those that do not, they struggle with fine-grained categorization of guilt types. Additionally, deep learning models and transformers, especially when utilizing contextual information from longer texts and a combination of titles and self-texts, show greater success in capturing the context of the text. Notably, the RoBERTa-base model achieved average F1 scores of 0.7599 for binary classification and 0.7394 for multiclass classification, outperforming all other experiments when combining the title and self-text.