Evaluating Latent Emotional Structures through Unsupervised Semantic Text Clustering

Abstract

Emotion analysis in textual data is an important topic in natural language processing, as emotions play a crucial role in understanding public opinion, psychological states, and dynamics of digital interaction. However, most existing studies rely heavily on supervised classification approaches based on predefined emotion labels, which may overlook latent semantic structures and emotional overlap inherent in natural language. This study aims to evaluate latent emotional structures in text using an unsupervised semantic clustering approach. The proposed method involves text preprocessing, feature representation using Term Frequency–Inverse Document Frequency (TF–IDF), dimensionality reduction through Singular Value Decomposition (SVD), and clustering using K-Means and Hierarchical Agglomerative algorithms. Both internal and post-hoc external evaluation metrics are employed to assess cluster quality and examine their correspondence with available emotion labels. The results indicate that K-Means clustering produces more compact and interpretable clusters than the hierarchical approach, while both methods reveal substantial emotional overlap across clusters. These findings suggest that emotional expressions in text exhibit a continuous semantic structure rather than discrete categorical boundaries. This study highlights the importance of unsupervised semantic clustering as an analytical tool for gaining deeper insight into latent emotional patterns in textual data.

Keywords

• Emotion Analysis
• Unsupervised Learning
• Semantic Clustering
• Text Mining
• Natural Language Processing