Aspect-Based Sentiment Analysis of Turkish School Reviews: A Comparative Study of Recurrent and Transformer Architectures

Aspect-based sentiment analysis (ABSA) of user-generated Turkish text remains underexplored, and no aspect-level resource exists for the education domain. We introduce a manually annotated dataset of 1533 school reviews from okul.com.tr, yielding 2620 (sentence, aspect) pairs over 19 aspect categories, imbalanced 2.3:1 toward positive polarity; a stratified 200-instance subset gave Cohen’s Kappa = 0.702. Given the aspect category, nine systems were compared under a 5-fold cross-validation protocol partitioned at the review level to prevent review-level leakage: four non-pre-trained baselines, two recurrent models (BiLSTM, BiGRU), and three fine-tuned transformers—the Turkish-specific BERTurk together with the multilingual mBERT and XLM-R, the three trained under identical hyperparameters, with class-imbalance handling applied where relevant. TF-IDF with a linear SVM reached a Macro-F1 of 0.836, exceeding both recurrent models, while all transformers exceeded every non-pre-trained system. In this dataset and experimental setting, BERTurk obtained the highest Macro-F1 (0.927) and accuracy (0.938), 0.091 above the strongest classical baseline. The advantage over mBERT concentrated on the minority class: F1-Negative exceeded mBERT’s by 0.034 points, roughly twice the accuracy gap. Its margin over XLM-R was small and not statistically established. The benefit of language-specific pre-training therefore appears to narrow as the multilingual model’s pre-training corpus grows.