Research And Application Of Multi-perspective Analysis And Generation Methods For News Content

With the rapid development of Internet technology,more and more people choose to browse news content on the Internet.The key to efficient use of online news media is how to quickly get the information you need from the vast amount of news information available.To make access to information more efficient,the first thing to consider is how to use existing resources to save users as much time as possible when looking for news.Secondly,it is necessary to extract the main information contained in the long news content,so that users can browse the high-quality news summary content.In addition to this,the content of news headlines is also crucial.Some unscrupulous media blindly seek clicks from users,and maliciously exaggerate and falsify news headlines.This has caused the news media to lose the credibility it once had and has brought many unsettling factors to social development.Based on the above reasons,this thesis uses deep learning methods to analyze and generate news text data from multiple perspectives.Three perspectives are investigating news recommendation list generation,news summary generation,and news headline generation.The details of the study are as follows.1)This thesis proposes a personalized news recommendation method based on news feature analysis and multi-view learning for generating news lists that match users’ interest preferences.The method is based on a multi-headed self-attentive mechanism to build two core components of the news encoder and user encoder.The features are extracted by analyzing the headline,body,and category information in the news content from multiple perspectives.First,the news encoder is used to encode and analyze the candidate’s news information to obtain news representation features.Then the user encoder is used to analyze the news information that the user has viewed to extract the user representation features.Finally,the news representation features and user representation features are used to generate a list of news recommendations that match the user’s interests and preferences.This saves users time in finding the news they need.2)This thesis proposes an extractive news summary generation method based on a pre-trained model for generating summary content containing news highlights.The method obtains sentence-level text feature vectors by extracting text features using a priori knowledge from the Ro BERTa pre-trained model.Then the sentence vectors are passed through a bi-directional long and short-term memory network layer and a multi-headed self-attentive layer to obtain document-level feature vectors.Finally,the prediction scores are output through the classification layer and the summary content is generated based on a combination of prediction scores.3)This thesis proposes a method for generating personalized news headlines that incorporates user characteristics to generate personalized headlines that match users’ reading interests and have authenticity.It can prevent the media from over-embellishing news headlines in pursuit of click-through rates.The method encodes the text vector by building an encoder based on the Fastformer model.In the decoding process,the user features extracted in the personalized news recommendation model are injected by the pointer generation network decoder.This influences the results of the generated headlines so that the content of the generated news headlines matches the user’s reading interests.Through experiments,it is proved that the three models proposed in this thesis for analyzing news text from three perspectives of news recommendation list generation,news summary generation,and news headline generation all achieve better evaluation results and the model performance is better than the baseline model.For the news recommendation list generation work,our results outperformed the baseline model by using AUC,MRR,NDCG,and other recommendation system metrics on the public news recommendation dataset.For the news summary generation work,the results were improved on the CNN/Daily Mail news summary public dataset using ROUGE-1,ROUGE-2,and ROUGE-L metrics.For the news headline generation work,the model outperformed other end-to-end generation models when tested on a manually authored test set.