Named entity recognition in Turkish celebrity news

This thesis focuses on named entity recognition in Turkish celebrity news. The aim is to evaluate the performance of deep learning-based models against the agglutinative structure of Turkish and the context-dependent nature of magazine language. For this purpose, a domain-specific dataset containing eleven entity types, including person, date, organization, location, and relationship, was manually annotated. Labeling was performed in both BIO and non-BIO formats, and the data was divided into training and test sets. Five transformer-based models, three designed specifically for Turkish and two multilinguals, were trained and analyzed using metrics such as accuracy, precision, recall, and F1 score, as well as confusion matrices and t-SNE visualizations. Comparisons revealed that the highest performance in both labeling systems was achieved by the mBERT model. In the identification of multi-word entities, the compatibility between the chosen labeling format and the structural features of the model proved to be a determining factor. In addition, LIME based explainability analysis was conducted to interpret the decision-making processes of the models. The results demonstrate that deep learning methods can be effectively applied to Turkish celebrity news and that high accuracy can be achieved when the appropriate model and labeling format alignment are ensured.