Multi-Label Emotion Classification Based on Plutchik's Wheel of Emotions
2nd International Symposium on AI-Driven Engineering Systems, ISADES 2026, Hybrid, Mbale, Uganda, 19 - 20 Haziran 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/isades69945.2026.11608219
- Basıldığı Şehir: Hybrid, Mbale
- Basıldığı Ülke: Uganda
- Anahtar Kelimeler: Emotion analysis, multi-label classification, Plutchik's Wheel of Emotions, rule-based derivation, two-stage emotion classification, XLM-RoBERTa
- Çanakkale Onsekiz Mart Üniversitesi Adresli: Evet
Özet
Traditional sentiment analysis approaches mostly evaluate texts across limited polarity categories such as positive, negative, and neutral; however, they remain insufficient in adequately reflecting the multidimensional, simultaneous, and intensity-based nature of human emotions. To address this limitation, this study proposes a fine-grained multi-label emotion classification framework that integrates Plutchik's Wheel of Emotions theory with the contextual and multilingual representation capabilities of the XLM-RoBERTa model. In the proposed pipeline, the machine learning models are trained exclusively to predict the eight primary emotions. The eight secondary emotions are not directly predicted by the models; instead, they are assigned through a rule-based matching process based on the co-occurrence of the primary emotions, following Plutchik's dyadic theory. This approach avoids conflicting secondary-emotion predictions and supports theoretical consistency without adding structural complexity. The same multilingual XLM-RoBERTa architecture was used for all three datasets: the Turkish Duygu-Turk dataset and the English GoEmotions and SemEval-2018 datasets. Duygu-Turk was evaluated using a two-stage protocol consisting of 5-Fold Cross-Validation for model selection and an independent heldout test set for final evaluation, while the official benchmark splits were used for GoEmotions and SemEval-2018. Experimental evaluations show that the XLM-RoBERTa architecture outperforms TF-IDF-based SVM and Random Forest baseline models. Evaluated only on the primary emotions, the model achieved Macro F1 scores of 98.55% on Duygu-Turk, 60.50% on the highly imbalanced GoEmotions dataset, and 46.33% on the noisy SemEval-2018 dataset. These results indicate that the proposed framework provides a theoretically grounded and cross-lingual approach for extracting complex emotional states from both structured corpora and noisy social media texts.