Research And Implementation Of Entity Recognition And Relation Extraction Based On Deep Learning

With the advent of the era of big data,information extraction technology has become an important tool to obtain key information from massive unstructured text.At the same time,in recent years,deep learning has been widely concerned and applied in the research field of natural language processing.In the big data environment,it can assist information extraction technology to achieve the goal effectively.However,in most specific fields(such as the medical field,the biological field,etc.),the application of deep learning method to extract information are facing the dilemma of sparse annotation data and poor model generalization effect.How to alleviate this dilemma in the case of insufficient training data has become a research hotspot.Therefore,this paper mainly explores the solution of information extraction using deep learning method in a small amount of labeled data environment from two aspects of named entity recognition and relation extraction.Based on the above problems and solutions,the work of this paper includes the following four aspects.First of all,for the task of Chinese named entity recognition,on the one hand,a new word discovery algorithm based on statistics and a new word filtering strategy based on word2vec algorithm are designed to fully discover the professional words in the specific domain.On the other hand,this paper proposes a new named entity recognition algorithm based on deep learning,which integrates the segmentation information into the lattice cells of the bidirectional short and long term memory network by using the results of the new word discovery algorithm,and introduces the attention mechanism to automatically pay attention to the important features,to improve the accuracy of the Chinese named entity recognition algorithm.Secondly,in order to achieve the goal of named entity recognition based on deep learning on the premise of reducing the amount of data annotation,this paper constructs a named entity recognition framework with active learning.From the perspective of more efficient use of valuable samples in the training set,this paper designs an iterative training process of named entity recognition,to realize the training of a high-performance model with low labeling cost.Then,based on the idea of transfer learning,with the help of domain(source domain)annotation data to alleviate the problem of sparse data in the current training domain(target domain),this paper proposes an algorithm of relation extraction based on adversarial learning.Taking convolution neural network as the basic network structure,and using the domain adaptive method,we add multiple adversarial network layers and a gradient inversion layer to the neural network,and use the discriminator and generator to achieve the feature alignment between the source domain and the target domain,while ensuring the alignment between the two domain categories.In addition,weight mechanism is introduced into the neural network to control the negative migration effect caused by outliers and negative samples,which effectively improves the accuracy of the relation extraction algorithm in the case of insufficient training data.Finally,based on the proposed algorithm of named entity recognition and relation extraction,this paper applies them to the information extraction system for verification.Using this system,users can obtain entity information and relation information between entities from unstructured text easily and quickly.In this paper,the requirements of the system are analyzed comprehensively,and based on this,the system scheme and the detailed realization scheme of each function module are designed,and the system functions are realized according to the design scheme.