Scholarlink

For gas emergency preparedness and response, emergency responsiveness has been enhanced through the establishment of multi-departmental coordination mechanisms and the deployment of intelligent sensing technologies. However, existing named entity recognition (NER) models suffer from limitations such as inadequate deep semantic representation and weak context modeling when processing accident case texts. These issues result in fragmented and incomplete recognition outputs, thereby compromising the real-time performance and accuracy of emergency response efforts. Therefore, this paper proposes a NER (BLSA-CRF) that incorporates multi-head local attention and hierarchical learning rate strategies. The model introduces a multi-head local self-attention (MH-SLA) mechanism to enhance local contextual semantic understanding. In addition, a full-module layer-wise learning rate strategy is adopted to set hierarchical learning rate configurations, which effectively improves the completeness of extracted information and enhances real-time emergency response performance. When evaluated on the dataset, the model achieved precision, recall, and F1-score values of 0.97, 0.98, and 0.97, respectively. Specifically, the precision outperformed that of the traditional model by 0.08, while both recall and F1-score showed improvements of 0.09. The results demonstrate the model's capability to extract relevant entities from gas accident texts, facilitating knowledge graph construction and improving domain knowledge application.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

BLSA-CRF: Named Entity Recognition Model with Multi-Head Local Attention and Layer-Wise Learning Rate Decay

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ning Li,
  • Qian Wang,
  • Anying Chai,
  • Chenyang Guo,
  • Lei Wang,
  • Enqiu He,
  • Junfeng Bai,
  • Jian Wang

摘要

For gas emergency preparedness and response, emergency responsiveness has been enhanced through the establishment of multi-departmental coordination mechanisms and the deployment of intelligent sensing technologies. However, existing named entity recognition (NER) models suffer from limitations such as inadequate deep semantic representation and weak context modeling when processing accident case texts. These issues result in fragmented and incomplete recognition outputs, thereby compromising the real-time performance and accuracy of emergency response efforts. Therefore, this paper proposes a NER (BLSA-CRF) that incorporates multi-head local attention and hierarchical learning rate strategies. The model introduces a multi-head local self-attention (MH-SLA) mechanism to enhance local contextual semantic understanding. In addition, a full-module layer-wise learning rate strategy is adopted to set hierarchical learning rate configurations, which effectively improves the completeness of extracted information and enhances real-time emergency response performance. When evaluated on the dataset, the model achieved precision, recall, and F1-score values of 0.97, 0.98, and 0.97, respectively. Specifically, the precision outperformed that of the traditional model by 0.08, while both recall and F1-score showed improvements of 0.09. The results demonstrate the model's capability to extract relevant entities from gas accident texts, facilitating knowledge graph construction and improving domain knowledge application.