Abstract: For solving the problems of the current fire information system, such as the complexity of information sources and the dependence of classification and interpretation on expert experience, this paper presented an automatic text summarization method to realize the auxiliary analysis of fire rescue text. On the basis of optimizing the fire rescue information dataset, the method contained an adopted pre-training BERT_WWM model, for extracting the word vector representation with context semantics, and used the Transformer to extract the summary sentence, so as to further improve the effect of fire information summary extraction. Through experiments in ROUGE-1, ROUGE-2 and ROUGE-L, our BERT_WWM + Transformer method was slightly improved comparing with other existed methods. Even the subjective evaluation could partially prove the purpose of extracting key information from texts, and showed that our method supported an available automation tools for intelligence analysis.

Key words: fire rescue intelligence, BERT, summary extraction, autoencoder