地理与地理信息科学

大数据时代海量社交媒体文本中蕴含的灾害信息是防灾减灾研究和应用的重要资源。该文针对灾害信息学领域面临的涉灾数据获取困难及信息抽取精度受限等问题，以地质灾害为研究对象，设计了一种“数据获取—数据识别—信息抽取”的方法框架，实现涉灾信息数据的准确自动识别，并进一步提升灾害信息抽取精度，从而有效获取地质灾害信息。具体而言，在数据获取方面，通过主题爬虫与SimHash算法实现数据采集和去重清洗，解决数据冗余问题；在数据识别方面，选取支持向量机分类模型对微博灾害短文本进行自动分类识别，其分类精度（大于90%）优于其他分类器模型，能准确识别灾害短文本数据；在信息抽取方面，构建地质灾害领域实体识别语料库并引入预训练模型和对抗训练层对网络模型进行优化改进，设计基于RBC-ADV的灾害信息识别模型，进一步提升地质灾害实体信息识别效果，实验表明，该模型在测试集上的表现优于当前主流模型。最后，为实现灾情信息的应用，针对地质灾害时空信息模糊性问题设计相应的推理补全方法予以解决，同时以“1·22云南镇雄山体滑坡”为实验案例，验证了RBC-ADV模型有效可行，能为灾害管理部门感知灾情信息、做出救援决策提供更全面的数据和信息支撑。

In the era of big data,disaster-related information embedded in massive social media texts serves as a valuable resource for both research and applications in disaster prevention and mitigation.Addressing the challenges of acquiring disaster-related data and improving the accuracy of information extraction in the field of disaster informatics,this study focuses on geological hazards and proposes a methodological framework entitled "data acquisition-data identification-information extraction" to enable accurate and automatic recognition of disaster-related information while enhancing the precision of disaster information extraction,thereby facilitating effective access to geological disaster data.Specifically,for data acquisition,the subject crawler combined with the SimHash algorithm is employed to collect,deduplicate and clean data,mitigating issues of redundancy.For data identification,a support vector machine(SVM)classification model is adopted to automatically categorize disaster-related short texts from Weibo.Comparative experiments demonstrate that the model achieves a classification accuracy of over 90%,outperforming other classifiers and enabling reliable filtering of disaster-related short texts.Regarding information extraction,a corpus for named entity recognition in the domain of geological disasters is constructed,and a disaster information recognition model based on RBC-ADV is designed.The model is further optimized by incorporating apre-training model and an adversarial training layer,which significantly improves the recognition performance for geological disaster entities.Experimental results show that the proposed model surpasses current mainstream models on the test set.Finally,to support practical application of disaster information,a reasoning and completion method is designed to address the ambiguity of spatiotemporal information in geological disaster reports.A case study of the "January 22landslide in Zhenxiong,Yunnan" is conducted to validate the effectiveness and feasibility of the proposed social media text mining method.The results confirm that the method can provide comprehensive data and informational support for disaster agency in situational awareness and rescue decision-making.