Typhoons cause major losses to human life and property each year in the Northwestern Pacific region. How to quickly collect information and make reasonable responses is an urgent problem faced by disaster relief departments. Crowd sourcing and citizen observation has been an effective method to obtain disaster information, among which social media, in particular represented by Twitter,¹ Facebook,² micro-blog data,³ etc., provide near real-time information during the disaster period. By making full use of the dynamic information collected by social media, the disaster relief department can get timely information about the disaster events and people's responses to them. Research has been done on the mining of disaster information based on social media data. Evidence shows that people's behavior is greatly influenced by social media when disasters occur.⁴ A study commissioned by the American Red Cross⁵ found that more than half of the respondents believed that government agencies should monitor social media to acquire timely and effective disaster information. As to how to use social media data to mine valuable disaster information, Chae J et al.⁶ used Twitter data for hurricane disaster analysis, and the results provided support for government departments' policy decision-making. Some studies^7,8 built disaster event classifiers based on microblog data for disaster event identification, which detected disasters through citizen observation. In addition, achievements have been made in the spatio-temporal analysis of disaster,^9,10 the characteristics of disaster social responses,¹¹ and the prediction simulation of disaster trends,^12,13 etc., which greatly improved the efficiency of disaster relief.

Collecting useful information for disaster events from social media is quite time-consuming and complicated due to unstructured expression. Although some social media platforms provide the API (Application Program Interface) for public information access, they also set restrictions to limit the information we can acquire. For example, we can't get the micro-blog information that relates to a specific disaster event; nor can we get the micro-blog information on a specified historical period directly through API. In other words, the API of these platforms does not provide corresponding retrieval functions, which undoubtedly increases the workload of subsequent data processing. Therefore, in our research project, we develop a toolkit to automatically harvest and process social media-based disaster information. We use the toolkit to generate a typhoon disaster dataset for 2017 based on several social media platforms. The dataset is mainly composed of text data that come from "Sina-Weibo" microblogs, "WeChat" Subscription and "Baidu" news. Figure 1 shows typhoon disaster data from "Sina-Weibo". The data contain textual descriptions and pictures of the disasters, as well as the time and location of data upload. It provides data support for the disaster relief departments to understand the timely progress of the disaster.

Data fields for "Sina-Weibo" includes ID, keyword, province, city, content, picture, location, release time, platform, number of forwards, comments, number of likes, as shown in Table 4. Each column has a limit of no more than 140 characters. The topics of the dataset include property loss, traffic impact, casualties, power supply, communication impact, rescue arrangements, response measures, and public attitudes toward the typhoon, among others.

Data fields for "Baidu" news include ID, title, link, source, release time, and keyword, as shown in Table 5. The fields for "WeChat" Subscription include ID, title, content, source, release time, and keyword, as shown in Table 6. The themes of the data include typhoon tracks, disaster loss statistics, government announcements, emergency measures, etc.

Keywords related to the designated typhoon event were diversified and optimized to ensure maximum retrieval of related information from each social media platform. After data collection was completed, we manually checked the validity of the data, and removed incomplete entries as well as entries irrelevant to the typhoon disaster. In addition, we established a database index system to avoid duplicate data. For disaster classification, three colleagues were arranged to classify these original data to ensure the accuracy of the final classification results. Prior to this, classification standards had been set up to minimize possible discrepancies. Finally, we randomly sampled 500 data entries from each platform and found an accuracy rate of nearly 100%.

To our knowledge, there were no social media-based datasets for these typhoons before, and our dataset effectively fills up this gap. The data in our dataset can be analyzed to meet different needs of disaster research. For example, the disaster loss data presented here can be re-classified into different categories to support real-time evaluations of disaster losses. The data can also be used for further analysis of typhoon disasters such as victims’ sentiment analysis in the typhoon area, the extraction of buzzwords during typhoon transits, etc. In follow-up studies, we have used the texts in this dataset to train the corpus for automatic identification of typhoon disaster information, which achieved satisfactory results.