Abstract: Abstract: The existing named entity recognition models based on BiLSTM are difficult to fully understand the global semantics of text and capture their complex relationships between entities. In this regard, a named entity recognition model based on global information fusion and multi-dimensional relation perception was proposed. Firstly, BERT was used to obtain the vector representation of the input sequence, and BiLSTM was used to further learn their context information. Then, a global information fusion mechanism composed of gradient stabilization layer and feature fusion module was proposed. The gradient stabilization layer enabled the model to maintain stable gradient propagation and update the representation of the optimized input sequence. The feature fusion module integrated the forward and backward representations of BiLSTM to obtain more comprehensive feature representation. Finally, a multi-dimensional relation perception structure was constructed to learn correlations between words in different subspaces in order to capture complex entity relationships in documents. In addition, the adaptive focus loss function was used to dynamically adjust the weights of different entity types to improve the recognition performance for minority entities. Experiments were conducted on 7 public datasets for the proposed model and 11 baseline models. The results show that F1 scores of the proposed model are higher than those of the comparison models, and the comprehensive performance is effectively improved.

Key words: Named entity recognition, global information fusion mechanism, gradient stabilization layer, multi-dimensional relationship perception, adaptive focus loss

摘要： 摘 要: 现有基于BiLSTM的命名实体识别模型难以全面理解文本的整体语义以及捕捉复杂的实体关系。对此，本文提出一种基于全域信息融合和多维关系感知的命名实体识别模型。首先，通过BERT获取输入序列的向量表示，并结合BiLSTM进一步学习输入序列的上下文信息。然后，提出由梯度稳定层和特征融合模块组成的全域信息融合机制。前者使模型保持稳定的梯度传播并更新优化输入序列的表示；后者则融合BiLSTM的前后向表示获取更全面的特征表示。接着，构建多维关系感知结构来学习不同子空间单词的关联性，以捕获文档中复杂的实体关系。此外，使用自适应焦点损失函数为不同类别动态调整权重来提高模型对少数类实体的识别性能。在7个公开数据集上将本文模型和11个基线模型进行了实验验证，实验结果表明所提模型的F1值均优于对比模型，综合性能得到有效提升。

关键词: 命名实体识别, 全域信息融合机制, 梯度稳定层, 多维关系感知, 自适应焦点损失