BERT for Arabic Named Entity Recognition
Chaimae Azroumahli, El Younoussi Yacine, Maciej Rybinski, José Francisco Aldana-Montes · 2020
In the recent year, transfer learning and contextual word Embedding models have majorly improved the performance of several NLP tasks for many natural languages. In this paper, we tackle the Arabic NER task while exploiting the effectiveness of using a context-dependent and transformer based representational language model. We train Arabic transformer-based representational language models using BERT [1], and compare them to the pre-trained multi-lingual model by building and evaluating the resulted Arabic Named Entity Recognition systems. Our approach utilizes BERT as multi-layer bidirectional transformer encoder that helped in learning deep bidirectional Arabic word representations, by training and fine-tuning the pre-trained model to generate the most accurate NER system for our datasets. Our system achieved the accuracy of 0.91 which demonstrated that creating a language-specific model can improve the performance of a specific NLP task in comparison to already trained Multi-language model.