A BIDIRECTIONAL ENCODER-DECODER MODEL WITH ATTENTIONMECHANISM FOR NESTED NAMED ENTITY RECOGNITION

Named entity recognition is a fundamental task for several natural language processing applications. It consists in identifying mentions of named entities in a text, then classifying them according to predefined entity types. Most labeling methods for this task use a label to recognize flat named entities because they belong to a single entity type. Therefore, they cannot recognize named entities that belong to multiple entity types.In this work, we concatenated all the labels of a word of a named entity into a joint in order to recognize flat or nested named entities. Then, we proposed a bidirectional encoder-decoder model with attention mechanism that uses this joint label to fine-tune a pre-trained language model for named entity recognition.We experimented our method on GENIA (a nested named entity dataset) and on two flat named entity datasets: CoNLL-2003 and i2b2 2010. Using the BioBERT model, our method achieved an F1 score of 78.85% on the GENIA dataset, 93.22% and 87.51% on CoNLL-2003 and i2b2 2010 respectively. These results show that our method can effectively recognize flat named entities as well as nested named entities.