Expanded Entity Coverage and Machine-Annotated Pre-Training for Urdu Named Entity Recognition

Named Entity Recognition (NER) for low-resource languages remains challenging due to limited annotated corpora and complex linguistic characteristics. Urdu is a morphologically rich Indo-Aryan language written in a cursive right-to-left script, increases these challenges in contrast with high-resource languages in NER performance. This paper presents a data-efficient framework for Urdu NER that combines large-scale machine-generated annotations with high-quality human-annotated data to mitigate annotation scarcity. Firstly, a new gold-standard human-annotated Urdu NER corpus is built comprising 49,040 tokens and 5,839 named entities across 13 fine-grained entity categories, including newly introduced types such as Sports, Food, and Color. To complement this dataset, a large machine-annotated corpus was created using a bootstrapped ensemble of Conditional Random Field models with confidence-based filtering. A two-stage training strategy in which multilingual transformer models, Multilingual BERT (mBERT) and XLM-RoBERTa (XLM-R) are pre-trained on the machine-annotated corpus and subsequently fine-tuned on the human-annotated data. Experimental results demonstrate that machine-annotation pre-training consistently improves NER performance for both models, yielding micro-averaged F1-score gains from 0.84 to 0.86 for mBERT and from 0.86 to 0.88 for XLM-R. Detailed per-class and confusion matrix analyses further show notable improvements for low-frequency and specialized entity types. The findings confirm that integrating weakly supervised machine annotation with multilingual transformer-based learning provides a practical and scalable solution for improving NER in Urdu and other low-resource languages. Keywords: NER, machine-annotated corpus, low-resource languages, Urdu NER corpus