Research And Implementation Of Nested Named Entity Recognition Based On Graph Attention Network

The rapid development of Internet technology has promoted the exchange and dissemination of information,but a large amount of unstructured text data in cyberspace affects the efficiency and quality of Internet applications such as Search and Reasoning.Knowledge graph-based Search and Reasoning technology can effectively improve the efficiency and performance of Internet Search.Nested Named Entity Recognition is one of the subtasks of named entity recognition in knowledge graphs.It aims at identifying entities nested in named entities.It can effectively improve the quality of named entity recognition and achieve the goal of building a comprehensive,correct and real knowledge graph.Graph neural networks can fully obtain contextual information when modeling unstructured text.In homogeneous graph neural networks,the main problems are the difficulty in boundary detection,high computational complexity,and difficulty in processing heterogeneous information.In the heterogeneous graph neural network,the main problem is that it is difficult to determine the central node and constrain the sampling domain of neighbor nodes.This work proposes nested named entity recognition methods based on the homogeneous graph attention mechanism and the heterogeneous graph attention mechanism to solve the problems of the graph neural network in the nested entity recognition task.At the same time,this work implements a nested entity recognition system based on the above two methods.The system also uses the Conditional Random Field and BERT for training and inference,and compares the evaluation indicators and inference time of the four models,which intuitively shows this paper.Experimental results straightly show the superior performance of the proposed model.Firstly,this work contains a nested named entity recognition method based on homogeneous graph attention mechanism to solve the problems of sequence segmentation,boundary representation and nested entity recognition in English.In this method,the context relationship between word and word,word and phrase,and word-phrase in the text sequence is calculated on the homogenous graph to complete the encoding of the nodes in the homogenous graph.In the end of this method,the bidirectional decoding module is designed to extract entities layer by layer.The nested named entity recognition method based on the homogeneous graph attention mechanism has natural defects in processing heterogeneous information,and at the same time,each calculation needs to load the whole graph into the computer,so the time and memory complexity is also high.Secondly,this work proposes a heterogeneous graph attention model that introduces Partof-Speech information of text to solve the problem of nested entity recognition.The method firstly initializes the heterogeneous graph nodes of the text through the Part-of-Speech information of the text,and then designs a node sampling algorithm based on the part-of-speech path to sample the neighbor nodes of the current node and update the current node with the attention mechanism.The bidirectional nested entity decoding layer described above performs nested entity decoding in a layer-wised manner.In the calculation process,both graph attention methods use the negative log-likelihood function to calculate the model loss.This method can model unstructured text on heterogeneous graphs,while sampling central nodes and updating neighbor nodes on small graphs can reduce the time and memory complexity of graph computation.Thirdly,based on the research contents of two kinds of graph attention networks,this work develops a nested named entity recognition system for biological texts,which can extract flat and nested entities in the biological field to improve quality of Knowledge Graph and performance on tasks such as Inference and Search.At the same time,the system also visually displays the evaluation indicators and running time of the model.The homogeneous graph attention model and heterogeneous graph attention model proposed in this work have improved Recall and F1-score compared with the same type of models,and the time complexity of the heterogeneous graph attention model is also lower than that of the same type of method.This work finally implements a nested named entity recognition system based on graph attention neural network.The system mainly includes nested named entity recognition method based on homogeneous graph attention mechanism and nested named entity recognition based on heterogeneous graph attention mechanism method.The system also shows the evaluation metrics and runtime of the model.The final results show that the method and system designed in this paper can effectively extract common entities and nested entities in the field of biology.