Abstract
Traditional Natural Language Processing (NLP) models for emotion classification often rely on explicit valenced words or local context, overlooking the underlying psychological drivers of human emotion. This paper proposes a framework to bridge this gap by integrating psychological constructs—specifically agency (the sense of control) and social interaction (the sense of connection)—into latent semantic spaces. Using a weakly supervised learning approach to augment BERT embeddings with these features, we demonstrate improved performance in predicting positive emotions and their temporal duration across multiple benchmark datasets.
Introduction
Human emotions are rarely isolated events; they are deeply tied to psychological states that are not always explicitly named in text. For instance, happiness might stem from achieving a personal goal (agency) or spending time with loved ones (social interaction). Current computational models often lack sensitivity to these latent "whys." This research investigates whether explicitly modeling these psychological constructs can better describe the semantic space of emotion and improve the predictive power of NLP classifiers.
Methodology
The researchers developed a generalized framework to map psychological constructs into the latent space of emotions using a three-step multi-modal approach:
- Weakly Supervised Labeling: Utilizing the CL-Aff HappyDB dataset—which contains labels for agency and social interaction—the authors trained a Multilayer Perceptron (MLP) classifier. This model was then used to "pseudo-label" other emotion datasets that lacked these specific psychological annotations.
- Feature Integration: A BERT-based architecture served as the backbone. To incorporate psychological insights, the predicted psychological labels (represented as a distribution) were concatenated with the 768-dimension vector produced by the BERT encoder.
- Multimodal Classification: This combined vector, representing both raw text embeddings and latent psychological features, was fed into a final classifier to predict the target emotion or its attributes.
Experimental Setup and Datasets
The framework was validated using several key benchmarks:
- CL-Aff HappyDB: 10,000 posts about happiness, used for training the initial agency and social interaction classifiers.
- Emotions in Text: A Kaggle dataset with over 20,000 observations.
- ISEAR (International Survey on Emotion Antecedents and Reactions): A high-quality dataset of ~7,600 observations collected by psychologists.
- SSEC (Stance Sentiment Emotion Corpus): A corpus of ~4,800 Twitter observations.
Key Results
Emotion Classification Performance
The inclusion of Agency (A) and Social interaction (S) features consistently enhanced classification accuracy. As shown in Table 2, the Bert Finetune + A + S model achieved the highest performance on the ISEAR dataset (97.66%) and the Emotions in Text dataset (84.61%). These results suggest that psychological theory provides tangible features that improve machine learning sensitivity beyond standard affective word recognition.
Temporal Dynamics and Duration
Beyond simple classification, the psychological features proved highly effective in predicting the duration of emotional states (e.g., whether happiness lasted "all day" or just "a few moments"). Within the HappyDB dataset, the integrated BERT model reached an accuracy of 72.52%, outperforming the standard BERT baseline and significantly exceeding the performance of simple MLP classifiers.
Discussion and Future Work
The findings indicate that agency and social connection are strong indicators of both the presence and the persistence of positive emotional states. By focusing on these constructs, models become less reliant on specific keywords and more attuned to the context of well-being.
As a student abstract, this work currently focuses on positive emotions. Future research will aim to generalize this framework to negative emotions (e.g., anger or sadness) and explore whether other constructs, such as empathy or coping mechanisms, provide similar performance boosts. Additionally, the authors plan to validate the internal consistency of their weakly supervised labels against direct human self-reports.
Summary updated