基于语义引导边界建模的嵌套命名实体识别模型OA

To address the insufficient modeling of internal structural information and relative boundary positions in nested named entity recognition,this paper proposed a structure-aware span modeling model based on a triaffine mechanism(SSMT).The token pair interaction module decoupled semantic and positional features between tokens and modeled their interactions,producing structure-aware attention scores that encode relative positional information.The sequence attention module focused on key regions within entities and enhanced internal structural representations.The fused structure-aware information then en-tered a dilated convolution module,which captured multi-scale contextual features.A triaffine attention constructed latent span representations,and the model concatenated them with the convolutional features for entity classification.Experiments on the ACE2004 and ACE2005 datasets achieve F1 scores of 88.64％and 87.30％,respectively.These results indicate that the pro-posed method is effective for nested named entity recognition.