Urdu Named Entity Recognition System Using Deep Learning Approaches.

The article focuses on advancing Named Entity Recognition (NER) for the Urdu language, which remains underdeveloped due to Urdu's morphological complexity and limited linguistic resources. It proposes a hybrid deep neural network architecture combining word embedding, character-level convolutional neural networks (CNN), and part-of-speech (POS) tagging to automatically learn features without manual engineering. Evaluated on four benchmark datasets—including a newly introduced Urdu tweets dataset (UTNER)—the model, particularly the bidirectional Gated Recurrent Unit (Bi-GRU) with CNN and Conditional Random Field (CRF) prediction layer, outperforms existing Urdu NER systems by improving the F1 score by 6.26%. The study also discusses challenges specific to Urdu NER, such as lack of capitalization, nested entities, spelling variations, and ambiguity, and highlights the benefits of incorporating character embeddings and POS features alongside word embeddings for enhanced recognition accuracy.