Table 3 Average test accuracy in five-fold cross validation of the proposed model and state-of-the-art methods on cross-sentence n-ary relation extraction. “-” denotes that the value is not provided herein. Full Parametrization (FULL) denote as each edge label is associated with a 2D weight matrix to be tuned in training. Type Embedding (EMBED) denote as each edge label to an embedding vector. K in the GCN models means that the preprocessed pruned trees include tokens up to distance K away from the dependency path in the lowest common ancestor subtree. *: significant at p<0.005
Method | Ternary | Binary | ||
|---|---|---|---|---|
Single | Cross | Single | Cross | |
Feature-based [31] | 74.7 | 77.7 | 73.9 | 75.2 |
LSTM-CNN [13] | 79.6 | 82.9 | 85.8 | 88.5 |
Graph LSTM-EMBED [5] | 76.5 | 80.6 | 74.3 | 76.5 |
Graph LSTM-FULL [5] | 77.9 | 80.7 | 75.6 | 76.7 |
Graph LSTM MULTITASK [5] | - | 82.0 | - | 78.5 |
GS GLSTM [12] | 82.3 | 85.5 | 85.4 | 85.6 |
GCN (K=0) [15] | 85.6 | 85.8 | 82.8 | 82.7 |
AGGCN [16] | 87.1 | 87.0 | 85.2 | 85.6 |
Bi-LSTM | 80.8 | 85.9 | 88.6 | 89.3 |
GNN | 83.0 | 86.6 | 88.7 | 88.6 |
Multihead attention | 81.5 | 87.1 | 89.7* | 90.6* |
With KG | 87.3* | 91.9* | - | - |