Research And Implementation Of Social Network Search Recommendation Algorithm Based On Bert

With the rapid development of the Internet industry,which gives help to the vertical development of the social network market.Compared with traditional deep learning and language representation models,BERT uses Transformer as the feature extractor,which has stronger feature extraction capabilities and can be used in a variety of different downstream tasks.Therefore,the paper is based on BERT for social network search recommendation.However,there are still many deficiencies in the.application of BERT in social network search scenarios:1)BERT doesn’t consider external knowledge in specific fields;2)Search scenarios are usually short texts,not a complete statements,and sometimes the user intent can’t be effectively analyzed;3)The BERT of the highest quality effect is too bulky and has too many parameters,so it is not right for search scenarios that have certain requirements on performance and resources.In view of the shortcomings of the current research work,the paper proposes corresponding solutions,The main content includes:(1)In the BERT encoding phase,the paper came up with a mask based on three levels of word,phrase and entity,and a coding method combining knowledge triples with BERT search sentences,and the negative sampling mechanism and the token layer enhances the model effect,keeping away from the noise problem caused by the introduction of information.(2)In order to make full use of the close semantic correlation between intent recognition and slot filling tasks,The two are jointly trained to improve the effectiveness of their respective tasks.(3)In order to compress the BERT model,the paper explored new distillation methods and text editing to enhance data set and model parameters while ensuring that the accuracy of the model is not overly affected.This paper integrates these methods to form a deep learning network model(Knowledge Embedded of BERT,KE-BERT)based on BERT and knowledge graph,and searches the data.After that,BERT combines TextCNN,TextRank and other technologies to recommend the searched data.The final experimental results verify the effectiveness of the method in social search recommendation scenarios.In addition,a code repository search recommendation system is designed,which carries KE-BERT model and hybrid recommendation algorithm to provide users with the search function of the repositories.