Research On Domain-related Named Entity Recognition Based On BISTM Feature Fusion

Natural Language Processing is a major study branch of the artificial intelligence project.This article is aimed at a key study point in Natural Language Processing: conducting a study on the task of Named-entity Recognition.The assignment is to recognize the phrases or words which have a certain semantics in a sentence,such as person name and place name.Nowadays,the task of Named-entity Recognition has taken all kinds of neural network as main models for recognizing.In recent years,research scholars gradually begin to consider integrating knowledge and the features of words into model recognition in the task of Named-entity Recognition,adding the knowledge of language habits into graph neural network architecture.Apart from this,they modify the neural networks and add various architectures,or utilize transfer learning to improve the ultimate effect of model recognition,large pre-training model and occurrence of its variant.Although those models are able to strengthen the impact,the speed of reasoning dramatically declines in the practical engineering application with the development of the models.Therefore,we come up with a neural network architecture in the article which is lightweight and can remarkably improve the model effect of the task of Named-entity Recognition at the same time.We create a neural network based on Bidirectional Long Short-Term Memory and a optimized model about Conditional Random Fields,put forward to add two extra features for multi-granularity embedding on the basis of phrase and word vector and combine the attention mechanism about word level and layer level to improve the model recognition effect.In addition,we add the physical representation of the text in the input layer of network architecture to achieve good results on entity recognition.The final recognition effect has been improved effectively by comparison.We divided into two parts of used data.On the aspect of model optimization,we adopted three classical corpus classifications in People's Daily,including the recognition about place name,person name and organization name.These three classified data are evenly distributed and representative.Eventually we demonstrated that F1-Score increased 0.7% after adding attention mechanism into the task of Named-entity Recognition,1.6% and 0.5% after the inclusion of the person name feature and the place name feature,and 1.9% after adding feature fusion.Overall,those features have played the role of optimization with 2.6% growth of F1-Score.We adopted the competition data of Chinese Named-entity Recognition from CCKS(China Conference on Knowledge Graph and Semantic Computing)on baidu.com,and compared the performance between two types of common neural network architecture in the task of Named-entity Recognition by screening the label of artificial and open source work.We also compared the performance between optimized Recurrent Neural Network and optimized Convolutional Neural Networks,and contrasted the recognition of these two network structures between Chinese person name and foreign person name with the same data.Finally,it is concluded that,in the task of Chinese Named-entity Recognition,the effect of Long Short-Term Neural Network Architecture is superior to that of the optimized Convolutional Neural Networks.And in the Chinese corpus,the F1-Score of Chinese person name recognition is a little more than that of foreign person name recognition with about 1%growth.