In order to tackle the challenges of uneven entity distribution and fuzzy boundaries in fire accident information extraction, a fire information extraction model RA-BMAF (residual attention-based multi-head fusion) based on adversarial training and multi-head attention mechanism was proposed. A total of 264 fire accident reports were used as samples for entity labeling, and fire text test data were constructed. The RoBERTa-wwn-ext model was employed to generate dynamic word vectors, and adversarial samples were generated by perturbing these vectors through adversarial training. These samples were then co-embedded with the original inputs for joint training, thereby alleviating the problem of entity distribution imbalance in fire samples. Additionally, in the feature extraction layer, the multi-head attention mechanism was incorporated to further optimize the weight distribution of the sequence features extracted by the BiGRU (bidirectional gated recurrent unit) model and to enhance focus on the fuzzy boundaries of entities. In the decoding layer, conditional random fields were used to predict the sequence features. The results show that the model achieved precision, recall, and F₁ scores of 83.10%, 91.13%, and 86.89%, respectively, outperforming the RoBERTa-wwn-ext-BiGRU-CRF (conditional random field) baseline model. It is concluded that the proposed approach significantly enhances the entity extraction process by addressing both entity imbalance and fuzzy boundaries in fire accident reports.