Font Size: a A A

Research And Application Of Named Entity Recognition Technology Based On Neural Network

Posted on:2021-11-11Degree:MasterType:Thesis
Country:ChinaCandidate:X H FangFull Text:PDF
GTID:2518306050455234Subject:Master of Engineering
Abstract/Summary:PDF Full Text Request
The emergence of smart TVs has broken through the limitations of traditional cable TVs.Through networking,you can achieve a series of multimedia functions such as watching network videos,listening to music,playing games,installing various software,KTV functions,online video calling,and watching the weather.However,it is very inconvenient to use a remote control to control many functions of the smart TV.To control the TV by voice will greatly improve the user's viewing experience.This paper is based on the humanmachine voice interaction system of TCL's next-generation smart TV,and is dedicated to implementing the named entity recognition system.The main work of the paper is as follows:(1)By analyzing the named entity recognition system,according to the diversity of user needs,a scene text classification based on convolutional neural network is proposed to obtain business scenarios,which are applied to named entity recognition to improve the accuracy of model recognition.The original data obtained from the web crawler cannot be directly used for network training,and needs to be preprocessed.The scene text classification model designed in this paper includes an input layer,a feature extraction layer,a merge layer,and an output layer.The input layer includes both word vectors and byte vectors.In this way,enter the corresponding feature extraction layer to obtain the feature vector,the convolution layer and the pooling layer of the feature extraction layer alternate,the merge layer stitches the feature vector obtained by the word vector and the byte vector,and the output layer includes a fully connected layer and Soft Max layer to classify the feature vectors obtained from the merged layer.In order to verify the performance of the classification model,the classification accuracy rate obtained through comparative experiments in this paper reached 0.9991.(2)The business scenario is obtained through the scene text classification model,and the corresponding NER model is called according to the specific business scenario.In order to obtain the NER model,this paper proposes a composite framework based on Transformer network and recurrent neural network to realize named entity recognition.The NER model based on the composite framework includes input layer,Transformer layer,recurrent neural network layer,and output layer.The input layer mainly generates Embedding vectors by means of random coding and position coding stitching.The multi-head Self-Attention mechanism of the Transformer layer can be calculated in parallel,thereby speeding up the network training speed,and can learn sentence-level sequence representation.The memory characteristics of the cyclic iteration of the recurrent neural network layer make it well able to learn word-level context semantic information and improve the accuracy of the model.The output layer includes fully connected layer and CRF layer.The fully connected layer can map the results of the recurrent neural network layer to the dimension of the label dictionary.The CRF layer can improve the accuracy of named entity recognition by automatically learning the constraints of entity labels.In order to verify the performance of the composite framework on the NER model,this paper designed a comparative experiment on the three networks of the composite framework,Transformer network,and recurrent neural network.The experimental results prove that the composite framework proposed in this paper can guarantee the accuracy and training time.(3)According to the above scene text classification module and named entity recognition module,this paper designs and implements a named entity recognition system based on the Django framework.Through the demand analysis of the named entity recognition system for smart TVs,the demand modeling was carried out,and the specific business process of each use case was introduced through the use case diagram and use case description.The function of the whole system is decomposed into scene text classification module and named entity recognition module.The function design and the function of each function are explained for each module.In the design and implementation process,it is introduced through the flow chart,and the system implementation interface is shown.In order to verify the system function,the system test was carried out,the test environment and test data set used by the system were introduced,and the test cases were written according to the test results.The results showed that the functions designed by the requirements were all met.By entering several text sequences,the system can predict the named entities contained in it,and all the evaluation results are correct,indicating that the model can be used in practice.
Keywords/Search Tags:Smart TV, Named Entity Recognition, Scene Text Classification, Composite Network Framework, Convolutional Neural Network
PDF Full Text Request
Related items