Joint Extraction Of Named Entity Recognition And Entity Relationship Based On Neural Network

With the gradual development of network informationization,unstructured text information continues to grow,and how to effectively process a large number of unstructured text information has become a research hotspot.Information extraction technology has attracted extensive attention from scholars because of its ability to extract from unstructured text information.Among them,named entity recognition and entity relationship extraction are important topics in the field of information extraction.At present,the methods for solving this problem are mainly divided into two categories: tandem extraction and joint extraction.The tandem extraction method is performed by first identifying the entity and then extracting the entity relationship.This extraction method defines named entity identification and entity relationship extraction as two independent subtasks.Its advantage is that each module is independent,flexible and easy to implement.However,it ignores the intrinsic relationship between the two tasks,and the result of the named entity recognition through the tandem extraction method will directly affect the subsequent entity relationship extraction,which is prone to error accumulation.The joint extraction method integrates the named entity identification and the entity relationship into a model.When there is a strong internal relationship or dependency relationship between the named entity identification and the entity relationship extraction,the information between the two can be better integrated,and each The errors generated by the intermediate steps,which in turn improve the performance of the extracted model.By analyzing the research status of the joint extraction method of named entity recognition and entity relationship,this thesis analyzes and experiments several methods that are effective in the field of joint entity recognition and entity relationship joint extraction.According to the shortcomings of traditional joint extraction model,this thesis Propose the following work:1.In order to solve a series of problems brought by the tandem extraction method and avoid complex artificial feature engineering,this thesis constructs a joint extraction model based on neural network for joint object recognition and entity relationship extraction.The joint extraction model based on neural network will be named.Entity recognition and entity relationship extraction are integrated into a model,which can reduce the manual extraction feature while fully considering the internal connections.2.For the typical joint extraction model,it is necessary to perform NER first in training,and then perform pairwise matching according to NER prediction information,which is easy to cause information redundancy.This thesis proposes a labeling strategy to transform the joint extraction problem into a labeling problem,which avoids the generation of information redundancy to a certain extent.3.The research focus of this paper is to extract the triples composed of named entities and relationships between entities.This extraction method needs to model the triples.Therefore,this thesis builds LSTM-LSTM end-to-end based on the new marking strategy.With a joint extraction model,the end-to-end model can be modeled directly through the LSTM neural network without the need for complex feature engineering.In addition,add a biased objective function to the LSTM decoding layer to make the tags more compatible with the LSTM-LSTM end-to-end model.4.Based on the LSTM-LSTM-Bias model,the attention mechanism is introduced to analyze the correlation between the input and output of the model,so as to obtain more con textual semantic information.The experimental results show that the LSTM-Att-LSTM-bias joint extraction model can identify the named entities more accurately and predict the entity pair relationship.The experimental results verify the validity and accuracy of the proposed algorithm.