Multi-label Classification Of Policy Text Based On Graph Neural Network

With the development of the information age,governments at all levels across the country have established official government websites through the Internet to display the government affairs information,so that more people can learn about it in a timely manner.At the same time in the context of the epidemic,most policies are closely related to the development and survival of SMEs,such as industry planning prospects,SME assistance policies,innovation and entrepreneurship policies,etc.Although the Internet is gradually covering the whole country,considering that to obtain this information,you must first frequently browse the websites of government departments related to your own enterprise,and at the same time read the specific content,so the smaller the enterprise scale,the worse the ability to obtain such information may be.And the appearance of policy text information has jumped in quantity.Government rules are formulated and industry segmentation is becoming more standardized and detailed.Policy text information presents the characteristics of more labels,increased levels and more detailed classification.In order to accurately classify policy texts and facilitate enterprises to search for policies related to their own industries,this thesis studies the multi-label classification of policy texts.The classification of policy texts as a practical application of text classification tasks can be handled by different methods based on statistics or based on deep learning.In the past,many methods are to consider the context of the text to classify the text,but considering the multi-label task,the label and the label have a strong correlation,in the long-tail distribution of multi-label data,the tail label often appears with a certain type of head label,so this thesis chooses to use a graph-based neural network,a deep learning method,to try to dig deeper into the correlation between different policy texts labels,from a new perspective to solve the label labeling problem of policy text.In this thesis,two models are proposed,one is a multi-label inductive lightweight diagram neural network(MLGPC)model that solves the problem of multi-label label labeling of policy texts for practical production,and the other is an inductive lightweight diagram neural network(ILGCN)model that plays a theoretical verification role in MLGPC and can be used for single-label policy text classification.In view of the many shortcomings of the graph neural network that are not suitable for Chinese policy text classification tasks,this thesis proposes a new graph data structure construction method,a new model structure and a data preprocessing method for the characteristics of the Chinese policy text,processing the Chinese policy text sample used as input from the three aspects of title,keyword and text,and using the graph neural network to mine the correlation between the tag-semantic graph data,combined with the information carried by the text itself to jointly provide the required information for the final multi-label prediction task.Improve accuracy.At the same time,the model theory of multi-label inductive light graph neural network is theoretically verified by the derived inductive lightweight graph neural network,which explains the rationality of the research theory and proves the application possibility of single-label policy text classification not covered in this thesis.