Statistical Research Of Shallow Semantic Analysis Based On Deep Neural Network Model

Semantic parsing is the process of identifying the formal representation of semantics from unstructured textual data and transforming it into structured data.Shallow semantic analysis is composed of sub tasks which are decomposed by semantic analysis and are suitable for general purpose.This paper focuses on how to effectively transform unstructured textual data into structured data.Based on the intrinsic connection between the various sub tasks,a statistical study on the shallow semantic parsing as a whole is carried out using deep neural network model.Firstly,a theoretical analysis framework is constructed,which consists of three parts.First of all,from four levels the connotation of shallow semantic analysis is studied.In three cases,the sub tasks of shallow semantic parsing are formalized into sequence labelling problems.It is necessary for modeling according to the intrinsic connection.Secondly,in terms of distributed assumption and multidimensional scaling analysis,the intrinsic unity between three kinds of word vector model is proved,which lays the theoretical foundation for effectively using the word vector in shallow semantic parsing.Finally,from the linear time invariant system theory,This article demonstrates the applicability conditions of convolutional neural network model when applied to shallow semantic parsing,the theoretical advantages of the long short memory artificial neural network model?the attention mechanism and the global optimization mechanism are analyzed.On this basis,a deep neural network model suitable for shallow semantic analysis is proposed,and the model construction,model selection and model prediction of shallow semantic analysis are systematically designed.In terms of parameter estimation,the gradient formula of the proposed model is derived,and the convergence of the improved algorithm is proved.Secondly,the empirical analysis and application of the analysis framework are conducted from three aspects.First of all,with the use of large-scale Chinese Baike corpus,a class of vector model is studied.The results show that,using the overall test based on rules of analogy,the accuracy of the model is as high as 89.24%.Secondly,using a multi-level corpus,on the basis of comparative study on the prediction effect of model components,with the predicate argument structure analysis as an example,the effectiveness of intrinsic connection is investigated in the process of modeling,which is compared with those of other similar models.For Chinese corpus,compared with the traditional model,when modeling without using intrinsic connection,deep neural network model enhances the model prediction effect by 11.18%;when modeling with intrinsic connection,only use part of speech analysis and named entity recognition,without changing the basic structure of the model,the average prediction effect further increased by 1.12%.For English corpus,the prediction effect of the model based on intrinsic relevance is significantly better than other similar models.All the results show the significant effect of systematic design through deep neural network.Finally,two applicable cases of the analysis framework are given,and the practicability and effectiveness of the shallow semantic parsing based on deep neural network model are verified by using the government service hotline worksheet of Dongguan.