Research On Video Entity Link Based On Multi-Source Incomplete Data

At present,with the popularization of mobile terminals and the speeding up of the network,short videos have gradually become popular with major platforms and users.A considerable part of the short videos are made by intercepting film and television resources,or by editing and processing these traditional film and television resources.For traditional film and television resources,each video playing platform has gradually built a mature knowledge base through years of accumulation.However,for the emerging short video,due to the timeliness and huge data,it is impossible to build a one-to-one corresponding knowledge base.Therefore,in view of the lack of short video knowledge base,we should use entity link technology to link the short video and the traditional long video knowledge base,and use a lot of prior knowledge in the long video entity to analyze the short video content,assist in judging the user's preferences,so as to accurately push the relevant content for the user.This paper mainly studies the link service between short video and long video entities based on multi-source incomplete data sets.For each short video,the dataset contains its text title and cover image.For each long video entity,the dataset contains its text description and cover image.The main task of this paper is to build an appropriate deep learning model for entity disambiguation,and find the most relevant long video entity from the candidate entity set.However,the short video cover map,long video text description and long video cover map in the data set are incomplete to varying degrees,which poses a great challenge to the task.In view of the above problems,this paper mainly does the following work:(1)To solve the problem of entity link between new short video and traditional long video,a new task of video entity link based on multi-source incomplete data is proposed,and the corresponding data set is organized and published.(2)To solve the problem of entity link based on multi-source incomplete data,this paper proposes an entity link algorithm based on similarity feature vector.This method firstly encodes the different dimensions of input data separately,and then calculates the similarity vectors of any two-dimensional input information between short video and long video entity respectively by means of full connection layer,Attention mechanism,VSM algorithm and other methods.Finally,the model outputs the result based on the similarity feature vector.(3)To solve the problem of fusing multiple sets of incomplete feature vectors,this paper proposes a feature fusion algorithm FISAM based on semantic alignment and Maxpooling.FISAM algorithm can make full use of each feature vector when the feature vector is complete,and maintain a relatively high accuracy when the feature vector is incomplete.(4)In order to improve the robustness of the model,a new data loading algorithm DLD based on dropout is proposed.DLD algorithm makes the model maintain a stable accuracy rate when facing the test data with different degree of incomplete,which shows a strong robustness.In order to verify the effectiveness of the various methods proposed in this paper,this project constructed a number of test sets with different degrees of incomplete to test each model respectively.The experimental results show that the method proposed in this paper has strong robustness.When the input data is incomplete,the model can maintain a relatively stable accuracy and achieve good overall effect.