Research On Bayesian Model And Variational Reasoning Method For Text Analysis

Bayesian model for text analysis is a method for deep mining of textual information.The approximation process of Bayesian model mainly uses variational inference method and Markov chain Monte Carlo method for estimation of hidden variables.In recent years,researchers have conducted inference analysis and research on Bayesian models,and there are few comparative studies on approximation methods for text analysis.In this thesis,the approximation methods are discussed on the Dirichlet allocation model,and a new Bayesian model is constructed and compared with the traditional Bayesian model.For the comparison of approximation methods,this thesis briefly explains the mathematical theory of variational inference methods and Monte Carlo sampling methods by introducing the research on the development of the field of variational inference in recent years.Several mainstream approximate inference methods are described,including the core methods and concepts of drop-free sampling methods,slice-sampling methods,mean-field variational inference methods,integrable variational inference methods,and hybrid variational inference.In the Gaussian mixture model comparison experiments,a Gaussian mixture model is established and several approximation methods are constructed,and several approximation methods are compared based on effectiveness,accuracy and other indexes.In the comparative experiments on textual datasets,a Dirichlet allocation model is built and the prediction performance and inferred topic performance of several variational inference methods are clearly compared on three datasets using two scoring criteria.In this thesis,in order to solve the problem of single topic distribution when traditional Bayesian models deal with multi-class texts,a Dirichlet hybrid generative model for text multiclass problems is constructed based on the Dirichlet assignment model.The model overcomes the limitations of the traditional Bayesian model and has a powerful fitting ability for the multiclass parameter variables hidden in the data collection,which helps to understand the whole data collection.Since the Dirichlet mixture generative model is an extension of the Dirichlet allocation model,it has the good properties of the Dirichlet allocation model.Different Bayesian models are constructed in the experimental part and compared on real data sets using two metrics,accuracy and normalized mutual information,and the results show that the Dirichlet mixture model has better subject assignment ability than the single-peaked Dirichlet model.The results show that the approximation ability of the Markov chain Monte Carlo method is better than that of the variational inference method,and the variational inference method has more advantages in terms of time efficiency.The integrable variational inference method outperforms other variational inference methods in terms of approximation and time efficiency.The distributive power of the Dirichlet hybrid generative model is significantly better than that of the Dirichlet distributive model.