A Hybrid and Explainable Feature Fusion Framework for Multi-Class Emotion Classification in Turkish Texts

Multi-class emotion classification in Turkish texts remains challenging due to the language’s agglutinative structure, morphological richness, and semantic overlap among emotion categories. To overcome these challenges, this research develops a hybrid framework that integrates lexical TF-IDF features with contextual representations generated by the pre-trained BERTurk model. Unigram and bigram TF-IDF features are combined with contextual embeddings to form a unified representation for emotion classification. The evaluation was performed on a publicly available Turkish emotion dataset consisting of 29,727 samples belonging to 13 emotion categories. Following data preparation, the corpus was partitioned into training and testing subsets through a stratified sampling strategy. Multinomial Naive Bayes, Logistic Regression, and Linear Support Vector Machine were evaluated as baseline models. The proposed hybrid architecture employed feature-level fusion and Logistic Regression as the classifier. A fine-tuned BERTurk model was also included as a transformer-based benchmark. Experimental findings showed that the hybrid framework achieved substantially better classification performance than approaches relying solely on lexical representations. While the strongest baseline model, Linear SVM, achieved a Macro F1-score of 92.79%, the hybrid model reached 95.66%, corresponding to an improvement of 2.87 percentage points. Five-fold cross-validation and statistical significance testing confirmed the robustness of the observed performance gains. Although fine-tuned BERTurk achieved the highest Macro F1-score (96.17%), the performance gap between the transformer and hybrid models was limited to 0.51 percentage points. The findings indicate that integrating lexical and contextual representations provides an effective and computationally efficient alternative to full transformer fine-tuning for Turkish emotion classification.