Research On News Recommendation Method Based On Deep Semantic Mining

The increasingly prosperous Internet has gradually changed the way people find and read news.More and more news readers tend to read news online.However,the amount of data provided by the Internet far exceeds the range that users can accept and process,resulting in information overload problem.The recommendation system is an effective solution.It is widely used in Internet online services by helping users quickly locate the items they are interested in.News recommendation,as a part of the recommendation field,has become a current research hotspot.Important semantic features of news are usually hidden in text segments of different granularities.At present,most news recommendation methods usually ignore fine-grained modeling of news item features,and it is difficult to fully dig out potential semantic information.In addition,the user's interest is dynamically transferable,and some of the user's interest may last for a long time,or may be triggered by a specific context or temporary demand.However,most methods treat user sessions as a single sequence,ignoring the evolution and correlation of user interests between different sessions.In response to the above problems,this thesis conducts research on news recommendation and proposes a news recommendation method based on in-depth semantic mining.The main work is as follows:First,in the stage of recommendation recall,this thesis proposes a candidate recall method based on text semantic mining.This method firstly obtains word-level semantic features by extracting the entities contained in news items based on word vector representation.At the same time,the convolution neural network is used to obtain sentence-level semantic features,so that the representation of news title is learned from multiple semantic views.Then,based on the neural topic model to model the content of the news text,and design a multi-topic capture module,by adaptively aggregating similar topics into topic representation vectors,thereby identifying the user's implicit reading interest.Finally,apply the attention mechanism to news title and text topics to better model the user's potential preferences and realize the recall of candidate news items.Second,in the recommendation ranking stage,this thesis proposes a news recommendation method based on deep interest perception,which focuses on the user's serialized behavior pattern.First,on the basis of the recall stage,use convolutional neural network to complete the extraction of news features;then,in the interest-aware attention layer,use gated recurrent units and self-attention mechanisms to model user behavior sequences in different periods;and finally in the session-aware attention layer,the attention mechanism is used to learn the correlation degree of user interest between the current session and the historical session,and integrate user interests in different periods to complete the click-through rate prediction of candidate news items,and realize news recommendation.Finally,this thesis designs and builds a news recommendation prototype system based on deep semantic mining.The construction of the prototype system includes requirements analysis,overall design,detailed design and other steps.The overall structure of the prototype system,the functions of each module and the specific operation process are introduced in detail.Through the effect display,the feasibility and effectiveness of the news recommendation method based on deep semantic mining in the actual application scene are reflected.