Mining And Searching Of Association Relationships Of Scientific Papers Based On Deep Learning

Scientific paper data is different from Internet big data,but an academic resource entity.Therefore,it is of profound practical significance to study the search technology of scientific papers based on deep semantic features and the mining technology of scientific papers association based on deep correlation model.Based on knowledge and technologies such as pre-trained language models,attention mechanisms,graph convolutional neural networks,variational autoencoders,etc.,this paper uses deep learning methods to mine and search for associations in scientific paper data.The main work of this paper is as follows:(1)A method for learning semantic features of scientific papers is proposed.According to the domain characteristics of scientific paper data,domain-aware scientific paper BERT is proposed.On the basis of the BERT model,scientific paper data is added for training,giving full play to the data advantages in the field of scientific papers,and achieving the effect of self-adaptation in the field of scientific papers.In the process of model training,this paper adopts strategies such as data adaptation in the field of scientific papers,mixed precision acceleration,and modeling light-weight to speed up the convergence time of model training on the basis of ensuring the accuracy of the model.Experiments show that the proposed method can obtain effective feature representation,and the training time is faster and the deployment costs are lower.(2)There are certain known correlations between scientific papers,such as citation relationship,co-author relationship,etc.This paper studies the application of existing information to mine deep correlations between scientific papers.A mutual information constrained variational graph autoencoders is proposed,which constructs scientific papers into graphs according to the existing relationships,and obtains node representations through the proposed encoder.The learned nodes represent the construction of a certain relationship graph,and obtain deep-level scientific and technological paper associations.Experiments show that the proposed deep association model for mining associations of scientific papers achieves better performance in various tasks.(3)The essential characteristics of short search word text and long paper text in scientific paper search scenarios are not fully considered,and the structural information of the text will be lost under extreme length differences.Therefore,this paper proposes a deep semantic matching model for scientific papers searching.An undirected graph is constructed by extracting keywords from the abstracts of scientific papers,which can effectively preserve the structural information in long texts.Based on the keyword undirected graph of scientific papers,this paper-further proposes a multi-view graph matching network for scientific papers based on the attention mechanism to form an interactive undirected graph from multiple views.The graph convolution network and graph aggregation network are used to output the interaction vectors from multiple views,and the multilayer perceptron is used to output the final matching results.The experimental results show that the proposed model achieves better search performance.(4)A deep learning-based scientific paper association mining and search system is implemented.The system integrates the semantic feature learning module of scientific papers,the correlation mining module of scientific papers based on the deep correlation model,and the search module of scientific papers based on the deep semantic matching model.It mainly realizes the following functions:scientific paper data collection,semantic feature learning of scientific paper data,algorithm visualization component,scientific paper association mining and scientific paper search.The system has complete functions,simple operation and friendly interaction.