Address Entity Identification Based On Bidirectional LSTM And CRF Models

Named entity recognition is an important research direction of natural language processing,acting as a key technology for many intelligent systems.In recent years,neural networks have developed rapidly,and the method of named entity recognition based on deep learning has become a new research hotspot.Aiming at the application requirements of the positioning of the insured address in intelligent auto insurance telephone reporting system,this thesis studies how to identify and classify the entity of address names from the description of users on their position.At the same time,combined with the practical application background of this thesis,a set of training set generation and optimization scheme is proposed on the acquisition of training data.The main work of this thesis is summarized as follows:(1)Due to the fact that the address names have huge quantity,and the naming methods are in great mess,a model based on multi-layer neural network is proposed to identify and classify the address entities.The network uses the BERT word vector as the model input,and uses twolayer bidirectional LSTM and CRF models for feature extraction.At the same time,the prior knowledge based on the address library is imported into the deep learning model.Through the accurate matching of the existing address library entries,the preliminary recognition result of the training corpus is obtained,and the result is spliced with the text word vector as the input of the model,to guide the model toward the direction of prior knowledge.(2)In the acquisition of training data,a complete and practical training set generation and optimization scheme is proposed.In view of the problem that the existing corpus cannot cover enough address names and user expressions,the method of filling the user expression language template with the address name is adopted to generate the training corpus.In addition,because the corpus obtained by the scheme may not fully meet the labeling specification,several rounds of optimization on the training corpus are executed,so that the model can achieve better recognition results.(3)Model hyper-parameter contrast experiments,training corpus optimization experiments and multi-model comparison experiments are conducted,and the model is applied to practical engineering.Different learning rate,LSTM vector dimension,training corpus quantity and DropOut parameters are contrasted in the model hyper-parameter contrast experiments,aiming to find the hyper-parameters which are more suitable for the subsequent experiments;Several rounds of training corpus optimization experiments are conducted in the training corpus optimization experiments,aiming to explore the influence of training corpus quality and prior knowledge on the model's efficiency;The multi-model comparison experiments are conducted referring to many other named entity recognition schemes,to study the effects of different deep learning models in address recognition tasks.After obtaining the optimal deep learning model,the model is applied to practical engineering,which verifies the feasibility of this model.The experiment results show that the deep learning model proposed in this thesis can achieve the effect of F1 value exceeding 90% in the context of large quantity address entities and messy user expressions.In the real operation of the intelligent auto insurance telephone reporting system,the address recognition model designed in this thesis can effectively identify the address information from the user's expression,which plays an important role in the operation of the entire system.