Named Entity Recognition (NER) in educational texts using transformer models with X (formerly Twitter)

Named Entity Recognition (NER) plays a crucial role in Natural Language Processing by identifying and categorizing named entities in unstructured text into predefined classes. Most research focuses on high-resource languages, leaving Urdu particularly in the education domain largely underexplored. This gap is largely due to the lack of annotated datasets for educational content, which limits the ability of existing models to accurately identify entities in this domain. To fill this gap, this study makes four key contributions. First, a large-scale Urdu dataset is developed for the education domain, containing 599,007 carefully annotated named entities across 11 key classes. Second, a detailed annotation process and corresponding guidelines are established. Third, key linguistic challenges such as morphological complexity and ambiguity are identified and resolved during the annotation of formal Urdu texts. Finally, ten experiments are conducted using 5-fold cross-validation, involving three machine learning models, two deep learning models, and three transformer-based models to identify the best-performing approach for the said task. Based on the results, the proposed model (XLM-R) achieves a cross-validation score of 98%, outperforming classical methods such as Random Forest (89%) by 10.11%. To further evaluate the proposed model, ablation experiments using an 80–20 train–test split show that a frozen XLM-R encoder reaches 91% accuracy, while full fine-tuning attains 98%, confirming the importance of end-to-end training. This work not only fills a significant gap in Urdu NER research but also establishes a robust foundation for future exploration of educational texts in under-resourced languages.