Research on Knowledge Representation Methods for Fire Incidents Based on Knowledge Graphs

Sudden fire incidents have become one of the most frequent and common major disasters that threaten public safety and social development. The data released by authorities is mostly represented in text form, which provides limited and unclear descriptions of emergency response procedures. The low level of data structuring results in some data being unable to be effectively transformed into valuable information for reuse, leading to the wastage of data resources and hindering further processing as well as subsequent applications of case event information. In order to structure and process fire incident information, and enhance the value of information utilization for assisting decision-making, we focus on the fire incident data recorded in a state in the United States as the case data. Considering the characteristics of fire incident data, knowledge representation learning methods based on knowledge graphs such the TransE model and the TransH model are applied. By adopting the idea of distributed representation, the data is embedded into low-dimensional dense vectors, which can significantly improve computational efficiency, alleviate data sparsity issues, and meet the requirements of downstream applications. Finally, the effectiveness of the proposed research method is validated by evaluating the embedding effects of the knowledge representation process using the evaluation metrics of link prediction tasks. Our experimental results demonstrate that TransH significantly outperforms TransE on link prediction tasks.