Urdu Named Entity Recognition with Attention Bi-LSTM-CRF Model

Abstract

The named entity recognition (NER) task is a challenging problem in natural language processing (NLP), especially for languages with very few annotated corpora such as Urdu. In this paper we proposed an Attention-Bi-LSTM-CRF method and applied it to the MK-PUCIT Corpus which is the latest NER dataset available for the Urdu language. In addition to word-level embedding, we used an embedding-level focus mechanism. The output of the embedding layer was fed into a bidirectional-LSTM encoder unit, accompanied by another self-attention layer to boost the system’s accuracy. Our Attention-Bi-LSTM-CRF model demonstrated an F1-score of 92%. The cumulative findings of the experiments show that our approach outperforms existing methods, thus yielding a new UNER (Urdu Named Entity Recognition) state-of-the-art performance. © 2022, The Author(s), under exclusive license to Springer Nature Switzerland AG.