This study presents Duygu-Turk, a novel deep learning-based sentiment analysis framework specifically designed for the Turkish language which is characterized by its agglutinative and morphologically rich structure. Unlike conventional sentiment analysis models that rely on coarse polarity classification (positive, negative, neutral) and insufficient integration of Turkish-specific linguistic features, Duygu-Turk adopts a fine-grained classification approach based on Plutchik’s Wheel of Emotions. The model identifies eight primary emotions, eight secondary emotions, and varying degrees of emotional intensity. Additionally, a non-monotonic logic mechanism is integrated to detect conditional sentiments, allowing for more context-sensitive classification. To enhance linguistic coverage, the model leverages morpho-semantic features, idiomatic expressions, suffixes, and contrastive conjunctions unique to Turkish. A new sentiment corpus consisting of 136,000 annotated Turkish sentences was constructed to train and validate the model. Experimental evaluations demonstrate that Duygu-Turk significantly outperforms transformer-based models such as BERT, DistilBERT, and ELECTRA, achieving F1 scores of 0.99 for polarity classification and 0.90 for multi-class emotion classification. These results highlight the model’s potential as a robust and linguistically grounded solution for sentiment analysis in Turkish and other low-resource languages. 

Sentiment analysis, Natural language processing, LSTM, Turkish language