Geographical Weighted Spatio-temporal Analysis Of Typhoon-related Public Opinion Based On Semi-supervised Learning

Social media contains a large amount of disaster-related public opinion text data,which has a strong real-time capability and contains a large amount of public emotional information,and is an important way of disaster situation awareness.However,social media data,with the low density of information fragmentation and many irregular language,leads to high data analysis costs.And few researchers consider the complexity of emotions contained in each text and the spatial heterogeneity of social media users.Issues such as how to obtain effective disaster data,improve the efficiency of information extraction,and accurately reflect the time and space response of the public when the disaster comes,and other issues need to be resolved urgently.This article takes typhoon disaster as an example.In order to reduce the cost of disaster public opinion data analysis,semi-supervised learning is introduced to eliminate irrelevant data with less labeling cost and achieve efficient extraction of disaster-related information;In response to the complex emotions contained in the text,the previous single-label sentiment classification was transformed into a multi-label sentiment extraction task,and public sentiment was evaluated from more dimensions,so that more public sentiment details can be presented;Finally the typhoon disaster was carried out based on the spatial heterogeneity of Weibo users.Geographically weighted time and space analysis of public opinion,so as to more accurately reflect the public's mental health and time and space response when disasters come,and further assist in optimizing disaster reduction strategies.The main content of the paper is summarized as follows:(1)Design and implement a data acquisition strategy for the Sina Weibo.In order to reduce the cost of environment configuration,cluster deployment is carried out in the form of containerization.Take typhoon as an example to obtain possible relevant Weibo data from Jun 8,2020 to Sept 15,2020,and perform data preprocessing to provide data support for disaster-related public opinion spatio-temporal analysis.(2)A semi-supervised text classification method based on Mixup technology called STCM is proposed,which introduces semi-supervised learning into disasterrelated public opinion analysis,and achieves efficient extraction of disaster-related information with less labeling costs.Using the pre-trained language model as the carrier,the sample mixup technology is used to generate new training samples in the hidden space of the text representation,combining other data enhancement technologies and unlabeled text low-entropy label guessing algorithms to build a semi-supervised text classification model STCM,and the Stacking model fusion strategy is used to construct the model training program.This semi-supervised text classification method has been experimentally verified to provide model support for the efficient extraction of disasterrelated public opinion information.(3)Based on the semi-supervised text classification method,the part of the acquired data that is not actually related to the typhoon disaster is removed,the finegrained sentiment of Weibo users contained in the disaster public opinion is extracted,and the microblog check-in data of each provincial administrative region is combined for geographic weighted analysis.The experiment found that during the active period of typhoons " Maysak " and " Haishen ",the movement direction of the public opinion center was highly similar to that of the typhoon,and it moved quickly to the southeast of China after the typhoon disappeared;Northeast China showed far more surprised emotions than other regions.Among the negative emotions,the public showed more sadness due to the rainstorm and travel disruption,but less fear and anger.