Nested Named Entity Recognition Based On Fused Features

Nested Named Entity Recognition(NNER)is an important task in Natural Language Processing(NLP),which aims to extract nested but non-overlapping phrase fragments from a sentence and classify them into categories.Nested named entities are a special case of named entities,which means that there is a hierarchy between different entities,and there are one or more entities inside an entity.In the corpus of academic research and all kinds of texts in real life,this kind of sentences containing nested entity structure is very common.Nested entities have richer entity knowledge and semantic knowledge.Making full use of these contents can mine deeper semantics of sentences and better understand text information.The improvement of the performance of the nested named entity recognition task is of great significance to the downstream tasks that depend on the named entities.Nested named entity recognition based on deep learning is one of the research hotspots.In recent years,researchers have proposed various related solutions for its special structure,and achieved good recognition results.In order to further improve the prediction performance of the model,this thesis has carried out in-depth research.Aiming at how to make entities of different nesting levels share information interactively and how to enhance the prediction of entity boundary and entity category by the model,the following two studies are proposed:(1)Nested named entity recognition method based on flat tag enhancement.In order to model the dependency of entities in different levels,a flat tag for nested named entities is proposed,which contains the information of the underlying entities and some of the upper entities.The method uses an auxiliary task to predict the flat tags,and the main task fuses the prediction results in a weighted way,so that the entities at different levels can interact with each other and learn some entity formation rules to improve the recognition performance of nested entities while avoiding error propagation.The experimental results show that the proposed method achieves better performance than the benchmark model in nested named entity recognition.(2)Nested named entity recognition method based on gating mechanism and syntactic information.It is found that the structure of nested entities is highly similar to the syntactic structure of phrases.If the syntactic information can be integrated into the two-stage method of nested named entities,it will be very beneficial to entity prediction.Aiming at how to effectively extract the hierarchy,boundary and category information from syntactic information,a gating mechanism is proposed to extract the valuable information for prediction.Specifically,two different gating neural networks are used to extract syntactic information and integrate it into the semantic vector of the model to strengthen the boundary prediction subtask and the category prediction subtask respectively.According to the results verified,the performance of the proposed method outperforms the effect of the benchmark model.