Traditional named entity recognition (NER) models and datasets have focused on mentions of persons, locations, and organizations, gaining good performance in well-structured English texts. Additionally modern language models like BERT have further improved NER models to achieve new state-of-the-art results. However, more complex scenarios with domain-specific entity types, shorter texts and different languages reduce the performance of NER models.
In order to create novel models which are able to improve NER performance in these set-ups, language models and NER models need to be adapted to take advantage of pre-trained models and incomplete structured knowledge.
More information can be found here.