Abstract:

[Purpose/Significance] Rich Internet data provide a multi-dimensional perspective for understanding emergencies, and multimodal emergency classification methods have emerged. However, the existing multimodal datasets of emergencies are not only scarce, but also lacking in diversity in categories, which is not enough to support related research, and greatly affects the progress of subsequent research. Compared with previous public datasets, the dataset constructed in this paper has richer categories and more improved modalities. This dataset solves the key gaps in the availability and diversity of multimodal datasets of emergencies. It not only expands the category range, but also provides more detailed classification in the natural disaster category, which is crucial for developing robust and accurate multimodal classification models. [Method/Process] An emergency event dataset (MEED) based on multimodal information was constructed, which contains data from five categories: accident disasters, public health, social security, natural disasters, and non-emergency events. The natural disaster data are divided into seven subcategories: geological disasters, biological disasters, drought disasters, marine disasters, meteorological disasters, earthquake disasters, and forest and grassland fires. [Results/Conclusions] The existing emergency classification methods were analyzed and validated on the emergency public dataset and MEED. The results showed that MEED helped improve the performance of multimodal models by more than 10% compared with the currently available emergency datasets. The results show that the improvement in model performance highlights the value of MEED in promoting emergency management and response research and applications. The dataset enables researchers and practitioners to better understand the complexity of emergencies and develop more effective prevention, mitigation, and response strategies. The improvement in model performance also shows that multimodal methods are a promising direction for analyzing emergency events because it leverages the advantages of different types of data to achieve higher accuracy and reliability in classification tasks. The creation of MEED is a major advancement in the field of emergency management, providing researchers with a valuable resource and potentially leading to the development of more sophisticated tools for responding to emergencies. However, the dataset still has certain limitations. Over time, the number of emergencies on the Internet continues to grow, which requires us to continuously update the dataset to adapt to new situations. The size of the dataset largely determines the performance of the classification model. The class imbalance problem of the emergency dataset constructed in this paper needs to be solved. In future research, we will continue to update and maintain the dataset in a timely manner to address these issues.

Key words: incidents, multimodal, dataset, deep learning, data acquisition, data annotations