HEEM, a Complex Model for Mining Emotions in Historical Text

Recently, emotions and their history have become a focus point for research in different academic fields. Traditional sentiment analysis approaches generally try to fit relatively simple emotion models (e.g., positive/negative emotion) to contemporary data. However, this is not sufficient for Digital Humanities scholars who are interested in research questions about changes in emotional expressions over time. Answering these questions requires more complex, historically accurate emotion models applied to historical data. The Historic Embodied Emotion Model (HEEM) was developed to study the relationship between body parts and emotional expressions in 17th and 18th century texts. This paper presents the HEEM emotion model and associated dataset from a technical perspective, and examines the performance of a multi-label text classification approach for predicting HEEM labels and labels from two simpler models (i.e., HEEM Emotion Clusters and the Positive/Negative model). The results show that labels in the complex model can be predicted with micro-averaged F₁ = 0.45, and macro-averaged F₁ = 0.24. Labels with fewer samples (<; 40) are not predicted. Overall performance on the simpler emotion models is significantly better, but for individual labels the effect is mixed. We demonstrate that a multi-label text classification approach to learning complex emotion models on historical data is feasible.