Enhancing Text Embeddings for Emotion Detection: A Study on Dimensionality Reduction and Lexicon Filtering

Abstract

Emotion detection in textual data is a crucial task in Natural language processing (NLP), yet standard word embeddings often fail to capture emotional nuances. This study explores an emotion-enrichment approach that refines text representations by integrating emotional information into word embeddings. In the study, two key challenges are primarily addressed: limitations of emotion lexicons, which may include ambiguous or misclassified words, and high-dimensional vector representations, which may increase computational complexity. To improve lexicon quality, which is an important data source in emotion enrichment studies, a filtering mechanism is introduced aiming to remove the words with inconsistent emotional associations, enhancing lexicon precision. Additionally, a sliding window-based dimensionality reduction method is applied to BERT embeddings to identify emotion-rich vector segments, reducing computational cost while preserving emotional information. Experiments are conducted in both English and Turkish to evaluate the impact of lexicon filtering and dimensionality reduction on emotion detection. Results show that filtering improves the accuracy of emotion-enriched representations, while sub-vector selection gives the possibility of finding more representative parts about emotional content. By focusing on emotion-relevant vector dimensions, the proposed method achieves superior performance compared to full-dimensional embeddings. This research contributes to multilingual emotion representation by refining lexicon-based enrichment strategies and optimizing embedding spaces for emotion detection. The findings highlight the importance of structured lexicon filtering and targeted dimensionality reduction in improving sentiment and emotion analysis models. © 2025 Elsevier B.V., All rights reserved.