Named Entity Recognition Based on Fusion of Different Channel and Spatial Information Features

At present, the leading method of planar and nested named entity recognition is to predict by word-word relationship, but there will be noise interference in multi-granularity dilated convolution. In the relationship prediction part, multiple Dropouts will discard important information and lead to model instability. In this paper, the improved convolutional block attention module (CBAM) is used to consider the relationship between different channels and spatial regions, and L2 regularization loss is added to the joint prediction layer to prevent overfitting without losing important features, which improves the overall model performance. Experiments show that the F1 score of this method is 0.28 %, 0.71 % and 0.92 % higher than that of the leading method on CoNLL2003, GENIA and Resume datasets, respectively, which proves its effectiveness.