Bert-Ner: A Transformer-Based Approach For Named Entity Recognition

The NER (“Named Entity Recognition”) technique picks up on named entities including people, places, companies etc., that are in the unstructured data in a new language processing technique. But, it arises a fundamental issue of the lack of localization that cannot be feasible. In spite of the achieved top performance of supervised NER methods in resource-rich domains, they usually too slow and not so accurate in domains with little labeled data. Here, the work present a new model known as BERT-NER which is an exceptionally deep learning approach built on the BERT (“Bidirectional Encoder Representations from Transformers”) architecture as its pre-trained language model. BERT-NER introduces the BERT installation to the NER mission making the context extraction be on the dot. BERT-NER fuses the BERT encoder and token classification layers together, optimized using cross-entropy loss. The experiment on the CoNLL-2003 dataset showed that BERT-NER could be an attention grabber as the scores were $\mathbf{9 3. 1 2 \%}$ precision, $\mathbf{9 4. 4 6 \%}$ recall, $\mathbf{9 3. 7 9 \%}$ F1 score, and 98.51% accuracy marking it as one of the field’s top results. The qualitative evaluation on the other hand, shows brilliance in systemic solutions across multiple domains. This project not only is about named entity recognition but focuses on creating broader domain in order to enhance the flexibility of the process. Through benefiting from transfer learning and initialized language models, BERT-NER increases the capability of the BERT model, in this manner extending a list of its applications in NER.