Research And Implementation Of News Recommendation System Based On Improved BERT And Gated Recurrent Unit Network

News has always been an important way for people to get information,understand current affairs and expand their horizons.With the rapid spread of mobile devices and Internet technology,news reading has shifted from newspapers and TV to online platforms.Facing the explosive growth of news production,a serious problem faced by users is information overload,so there is an urgent need for a means to filter information in a short period of time to obtain news content that meets users’ interests and needs.However,due to the difficulty of news text feature extraction,sparse data,and the susceptibility of user interests to drift,existing news recommendation systems can not well meet the growing needs of users.To address the above problems,this thesis implements a personalized news recommendation system from two strategies,using a feature projection algorithm based on a deep bidirectional encoder and a gated cyclic unit algorithm based on multi-feature fusion.This thesis mainly includes the following elements:(1)Aiming at the problems of irregular vocabulary,unclear semantics and sparse features in news subject articles,a feature projection algorithm based on deep bidirectional encoder(DBERT-FPNet)is proposed for news recommendation.The method firstly uses a crawler to obtain the news text.The method firstly uses crawlers to obtain news data and pre-trains a lightweight bidirectional encoding representation model to classify the obtained news data.Then,the processed data are fed into the DBERT-FPNet algorithm for feature extraction.The DBERT-FPNet model consists of three modules:word embedding,DBERT pre-training model and FPNet model.Among them,the word embedding module is mainly responsible for expressing words into static vectors,and then the DBERT pretraining model extracts the global semantic feature information;the FPNet model is responsible for purifying the text features finally extracted by the DBERT model combined with the feature projection method,thereby strengthening the classification effect.Finally,extensive experimental research is conducted based on four datasets of Toutiao,Sohu News,Sina News and Small Sina News.The results show that compared with the pre-trained BERT-Bidirectional Gated Recurrent Unit(BERT-BIGRU),the method proposed in this paper has improved accuracy on the four data sets.(2)To address the problem that current news recommendation algorithms do not sufficiently mine news content as well as users’ long-and short-term interests,a multi-feature fusion based gated recurrent unit news recommendation algorithm(MFFGRU)is proposed.algorithm(MFFGRU).The algorithm first obtains the potential topic distribution from the news body based on the implicit Dirichlet distribution,and then learns a unified news representation based on the textual contents of the news such as headline,abstract and body,as well as additional information such as explicit topics and potential topics.Finally,this thesis adopts Gated Recurrent Unit(GRU)to dynamically capture the news feature information over time to mine the sequential interest features of users,and also to model their long-and short-term interests.In addition,in order to explore users’ current concerns and stable preferences,the algorithm introduces an attention mechanism to achieve personalized recommendations.Experimental analysis found that when evaluating algorithm performance on Microsoft News Dataset(Microsoft News Dataset,MIND)and MIND-small,compared with the lightweight graph convolutional neural network,the area under the curve of the proposed model has been improved.(3)In order to help users quickly find personalized and high-quality news content and improve reading experience,this thesis designs and implements a deep learning-based news recommendation system based on the algorithm proposed above,which mainly contains four modules.Users can register,log in and manage their personal information through the user module,and browse news information in the news information module,and sort and filter according to different classification,time,hotness and other users can register,log in and manage their personal information through the user module,and browse news information in the news information module,sorting and filtering according to different categories,time,hotness and other factors.The news management module is mainly used for publishing,modifying and deleting news,and it is based on the recommendation algorithm proposed in this thesis,which can implement personalized recommendations based on users’ interests and preferences.In addition,users can also like and comment on news content through the user interaction module,which facilitates the analysis of user behavior at a later stage.The system can help users get the news they are interested in more conveniently,and also provide more accurate audience targeting for news publishers,which is an innovative system with wide application prospects.