| Abstrakt: |
Visualized knowledge representation can more effectively help the public gain knowledge about lung cancer prevention, diagnosis, treatment, and subsequent life. Therefore, this study collected articles on lung cancer from the well-known Web of Science database to analyze lung cancer literature, and the text data were published between 2016 and 2021. First, we used natural language processing to handle the collected text data, and then we used the latent Dirichlet allocation method to perform topic modeling and obtain the optimal topic numbers based on two coherence metrics for assigning the class of every article. Next, a PMI_2 weighted was proposed to build an initial weighted knowledge graph, and four graph neural network algorithms were used to train the initial weighted knowledge graph. In addition, we proposed a PMI_2 + link to improve the classification performance, and the additional links were obtained from the graph auto-encoder and graph convolutional network training. When the best classification performance has been obtained, these edge weights have a representative. For visualized knowledge representation, we used the Neo4j tool to display the nodes and edge weights for the final literature knowledge. The results show that the use of the proposed PMI_2 + link to build a weighted graph has a better classification performance. Further, the proposed PMI_2 + link can effectively reduce the number of edges on the knowledge graphs and avoid insufficient GPU memory. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink