Candidate Region Aware Nested Named Entity Recognition

Named entity recognition(NER)task aims to locate and classify entities into pre-defined semantic categories(e.g.,person names,organizations and locations)from unstructured text.NER is a fundamental task in natural language processing(NLP)and benefits many NLP applications,such as knowledge graph construction,entity relation extraction,entity linking,question answering system and so on.However,most of current works on NER ignore the nested entities which are common in many realistic applications.The extraction of nested entities can help to improve the performance of downstream tasks.This research proposes a neural multi-task model which is capable of recognizing nested entities and is named candidate region aware model(CRAM).To alleviate the heavy class imbalance in existing approaches and improve the model efficiency,this research divides nested NER task into two sub-tasks: the entity-token sequence labeling task and the candidate region classification task.Two modules are proposed for the two sub-tasks respectively.The entitytoken sequence labeling module figures out whether the tokens are entity-tokens in input sequence or not.Then,the candidate regions can be obtained and classified into corresponding entity categories with the candidate region classification module.Considering the correlations between two sub-tasks,the two modules are trained jointly with a multi-task learning mechanism,which is effective and efficient.To validate the effectiveness of our research,extensive experiments are conducted on several public datasets.The experimental results show that our proposed candidate region aware model can effectively alleviate the class imbalance problem in existing approaches and achieve better performance.Besides,under the same experimental environment,our CRAM model has the shortest training time among the compared neural network-based approaches.