Research On Domain Information Mining And Theme Evolution Of Scientific Papers

In recent years,with the increase of social investment in scientific research,the number of research results in various fields has increased significantly.Various research fields are developing rapidly,and different fields are cross-integrated,and there are certain dependencies between many research results.How to effectively mine the research field and theme evolution law of scientific papers has important research significance.Based on attention mechanism,graph convolutional network,graph attention network,hierarchical label classification network,temporal convolutional network and other deep learning technologies,this paper presents the feature representation of scientific paper data,uses the hierarchical multi-label classification model for scientific paper domain information,and conducts research on the evolution law of various topics in the research field.The field information mining and evolution law prediction system of scientific papers is established through micro-service technology.The main work done in this paper are as follows:(1)A method for data acquisition and feature extraction of scientific papers is proposed.Aiming at the feature representation of scientific papers,this paper proposes a semantic feature representation model for scientific texts that combines graph convolutional network and BERT deep semantic model.A heterogeneous network of papers and keywords,the initial vector representation of all nodes is obtained through BERT,and the semantic features of adjacent nodes are combined with the GCN network to extract spatial information,and the final output is used as the semantic representation of scientific papers,realizing the semantic feature representation learning of the paper.(2)An information mining model in the field of scientific papers based on graph convolutional networks is proposed.A paper often has multiple categories.For example,in the CLC classification method,there is a hierarchical relationship between categories,and the categories of the paper present a hierarchical structure,which is a typical multi-label hierarchical classification task.It poses two challenges to existing models:First,existing models do not capture the semantic relationships between papers well.Second,they neglect to model the hierarchy of labels.This paper proposes a hierarchical label attention model based on graph attention network,which exploits the co-occurrence of words to model the semantic relations of papers.Hierarchical multi-label classification of paper domains is achieved by using multiple linear layers to model category hierarchies and combining each hierarchy of labels through an attention mechanism.(3)A prediction model for the evolution law of scientific papers research topics based on spatial enhancement and dynamic graph convolutional networks is proposed.Since there are certain correlations between various research topics,looking at a single research topic in isolation cannot effectively explore the dependencies between these research topics.In order to simultaneously capture the spatial dependencies and temporal changes between research topics,a research topic trend prediction model based on spatial augmentation and dynamic graph convolutional networks is proposed.This model combines graph convolutional neural networks(GCN)and temporal convolutional networks.Network(TCN).GCN is used to learn the spatial representation of research topics,which uses spatial dependencies to enhance spatial features.TCN is used to learn the dynamic changes of research topic trends,which optimizes by calculating weighted loss according to temporal distance,realizing the evolution law of research topics excavation.(4)The domain information mining and theme evolution law analysis system of scientific papers is implemented.The system integrates a data processing module,a subject query module,and a research theme change trend module.Based on the Springboot framework,a visualization system is developed,and the ElasticSearch retrieval engine and Redis cache database are integrated,which effectively improves the performance of each module of the system.The Echarts visualization plug-in is used for visual display,providing retrieval of Chinese and English paper results and distribution analysis of paper subjects,research theme correlation map,research theme evolution law prediction and other functions.The functions of the system are complete,and the user interface is friendly to operate.