关键词: 地震应急, 自媒体, 语义分析, 危急度

Abstract: From the events of catastrophic natural disasters that have occurred in recent years, it can be found that social media platforms are increasingly becoming the most important and most convenient way for the general public to timely release and obtain information on disasters. The information obtained from such platforms contains a large amount of information in the form of texts, pictures, etc. that record the current situation of the disaster. And it also has characteristics of high efficiency and high spatial distribution to serve the rapid emergency after the earthquake. In this paper, we firstly make a statistical analysis of 32 689 pieces of historical disaster data acquired from 5 earthquakes with obvious characteristics, such as post-earthquake disaster events, user's expression habits and so on, and adopts cross-validation method. Then information classification system which includes seven first-level categories and more than 50 second-level categories is constructed. The information classification system and evaluation system of crisis degree for post-earthquake emergency response are constructed both using cross-validation method. The former is referred to the thought of existing classification basis and the experience knowledge of several emergency experts. Based on the five indicators of subject word, action word, degree word, time and position measurement, an evaluation system of critically with four levels of severity, moderate intensity, mildness and others was constructed. Considering the sparse features of self-media information and the large difference in the number of training sets, a naive Bayes model for information classification is trained based on the classification system and evaluation system. Its accuracy rate is 73.6%. At the same time, the classification method of feature fusion of machine learning model and semantic calculation model is used to evaluate the criticality of the disaster information. The accuracy rate of the evaluation model is 89.2%, higher than 85.2% of the semantic computing model and 77% of the naive Bayesian model. The evaluation model has combined the advantages of semantic computing method which can evaluate all index features with machine learning method which has high classification efficiency and accuracy. The thresholds for classification between mild and moderate intensity, moderate intensity and severe intensity were 15.2 and 27.39. The model realized in this paper can crawl, classify and evaluate the disaster information in the media in real time after an earthquake, and realizes mining of a small amount of critical and important information from the massive self-media information, thus, to assist in earthquake intensity rapid reporting and accurate rescue. Finally, taking the Jiuzhaigou earthquake on August 8, 2017 as an example, 17 432 pieces of data were crawled in real time within 48 hours after the earthquake. At the same time, based on ArcGIS, the mining information is visualized in time and space, and the availability of the data is evaluated from two perspectives of earthquake intensity quick reporting and accurate rescue after the earthquake. The disaster information of Jiuzhaigou County in the earthquake area is obviously more than that of the non-earthquake area in terms of quantity and emergency degree. The results show that the self-media information with high spatial distribution can effectively find the severer disaster grade area after the earthquake, shorten the time of earthquake intensity prediction, effectively classify and extract information, provide real-time information for precise rescue, and improve the efficiency of emergency response after the earthquake.

Key words: earthquake emergency, self-media, semantic analysis, criticality