Chinese medical named entity recognition based on zero-shot learning

Abstract

To solve the problem that existing named entity recognition models lacked of ability to deal with unseen classes, the zero-shot learning was proposed to be used in the task of Chinese medical named entity recognition. Zero-shot learning utilizes the description information of the entity's class to establish the connection between the entity and the class, and transfers information from the observed classes to the unseen target classes. The model proposed by this paper was mainly based on BERT, which was used to model the relationship between the entity and the description. Moreover, the static word embedding, which is as a supplementary information, is concatenated with the features obtained from the BERT to solve the problem that BERT is not suitable for a specific field. At the same time, Correlation Searchers are added between the transformers of BERT to search for the word information most relevant to the character, so as to solve the problem that the model cannot obtain complete word information with characters as the input unit. Experiments show that the model's recognition performance has been significantly improved after adding the static word embedding and word information.