Fig 1.
The overall network structure of MaskDGNets.
Where, ‘Bi-Indrnn+’ indicates an improved bidirectional independent recurrent neural network module. ‘WVL’ indicates the word vector layers. ‘CEL’ indicates the character embedding layer. ‘EM’ indicates the embedding module. ‘DGAM’ indicates the dynamic graph aggregation module. ‘MaskAM’ indicates the mask attention module. indicates word vector feature with word and characters, the sentence sequence features and fusion feature by mask attention module.
indicates the word feature and character feature. ⊗ indicates the feature matrix multiplication. BN(⋅) indicates the batch normalization operation. Weight indicates the weight matrix. Recurrent represents the recurrent unit. ReLu represents the activation function. AdPool(⋅) represents the adaptive pooling operation. © represents feature concatenation. Conv3×3(⋅) indicates 3 × 3 convolution operation.
indicates the graph nodes, where
indicates the event node in the sentence,
represents the basic attribute node of the event in the sequence S.
Table 1.
Experimental results of different event extraction methods on the DuEE and CCKS2020 datasets.
Table 2.
Ablation results of different module with our proposed MaskDGNets on the DuEE and CCKS2020 datasets.
Table 3.
Ablation results of reconstruction loss function with proposed MaskDGNets on the DuEE and CCKS2020 datasets.
Fig 2.
Validation loss of different event extraction methods.
Fig 3.
Efficiency of different event extraction methods.
‘FLOPs’ indicates Floating Point Operations Per Second. ‘Params’ indicates the parameter quantity.