利用卷积神经网络提取微博中的暴雨灾害信息

从社交媒体中挖掘灾害应急信息,能够有效帮助传统灾害管理获取实时、主题丰富的灾害信息,从而成为灾害应急管理的新手段。得益于深度学习在自动特征提取上的成就,本文研究了一种利用卷积神经网络对社交媒体中的灾害应急信息进行自动实时提取与分类的方法。首先,利用社交媒体数据和Word2vec模型,构建与灾害类事件相关的语料库并获得相应的词向量;其次,将词嵌入文本和相应的灾情类别作为卷积神经网络的输入,经过多分类学习得到分类模型,用以提取近实时灾害信息。以2012年“7.21北京特大暴雨”事件为案例,通过分类模型获得常见灾情类别的暴雨灾害社交媒体信息。该模型在测试集上的精度达到了90%以上,并且将模型运用于新爬取的2016年暴雨数据集上也得到了较好的表现,说明该模型在近实时自动提取灾害信息方面具有可行性。在对2012年分类结果进行时空分析结果表明,通过社交媒体获得的暴雨灾害主题信息符合灾害发展的规律,说明了利用深度学习提取社交媒体数据中的灾害应急信息的有效性和可行性,能够为实时灾害应急管理提供新的思路。

Extracting Rainstorm Disaster Information from Microblogs Using Convolutional Neural Network

Nowadays social media has played an increasingly significant role in disaster management, thanks to its real-time nature and location-based services. When a disaster happens, a large number of images and texts with temporal and geographic information quickly flood in the social media network. Complementary to the traditional disaster management, social media could provide a lot of dynamic, nearly real-time disaster information to researchers. Current studies place more emphasis on using machine learning to deal with social media disaster data. Yet, in many cases deep learning has a better performance in automatic feature extraction than the traditional machine learning, and it can be used to extract and classify disaster information from social media. This paper focused on a method of extracting the disaster information from social media data using Convolutional Neural Network (CNN). To obtain the word vector corresponding to social media texts, a corpus of disaster events by using social media data was trained by word2vec model. Then, the vectorized microblog sentences and their corresponding disaster categories were used as input to the multi-classification model, which is based on convolutional neural network. After training and optimization, we used this model to extract disaster information from a large number of social media data streams. For an experiment, we combined Sina Weibo API and web crawler, and got over twenty thousand microblog texts with the theme of "Beijing Heavy Rainstorm" happened in 2012. Besides the irrelevant texts, we divided the data into seven categories. The topic classification model of rainstorm disaster information was built and trained based on a small number of tagged Sina Weibo data. The experimental results achieved the F-value of over 80% and the precision of over 90%, proving the validity of applying the model to our dataset. Moreover, this model when used to classify the data on Beijing's rainstorm in 2016 newly crawled form Weibo also had a good performance. According to the different rainstorm emergency topics classified by model, we carried out the deep mining of time series and spatial features to detect the phases of disaster development. Through visualization and statistical analysis, it was found that the time series analysis of disaster was consistent with the development of actual disasters, indicating the effectiveness of the CNN-based method in monitoring Beijing rainstorm. The study shows that using deep learning to extract disaster emergency information from social media is effective and feasible, which provides a new approach to real-time disaster emergency management.