We are presenting an integrated framework for explainable Turkish text classification applied to the Internet news on brain drain and migration attitudes. We construct a 2,343-article migration discourse collection corpus, preprocessed via Unicode normalization, Turkish‐specific lemmatization, and sequence length pruning to 978 subword tokens. Four Transformer paradigms, encoder‐only (BERT), decoder‐only (GPT), masked-only decoder (MaskedDecoder), and encoder–decoder (T5), are re-fined-tuned with low‐rank adapters under an architecture‐preserving optimization schedule. A novel cluster consistency loss aligns Transformer hidden representations with class‐specific keyword clusters extracted by YAKE, RAKE, Transformer models with Sentence-Transformer/KeyBERT Pipeline, BERTopic, and LDA. Quantitative interpretability metrics, including head‐wise sparsity, layer‐wise Kullback–Leibler divergence, and SHAP‐based regional importance, reveal that lower layers gather broad evidence while upper layers concentrate on class‐defining tokens. Qualitative cross‐attention visualizations confirm a sharp semantic boundary at the token “başbakan,” with selected heads faithfully surfacing centroid keywords. The proposed system achieves an accuracy of 0.7952 and weighted F1 scores of 0.7932 on migration news classification, while delivering human‐intuitive explanations that satisfy both fidelity and transparency. Our work instead focuses on structured attention-based interpretability within Turkish discourse classification, aiming to balance model performance with linguistic transparency across societal domains like migration and brain drain.
Yıldırım et al. (Thu,) studied this question.