Fig. 2

Overview of our model. The Bi-LSTM first encodes each word by concatenating word and position embeddings, followed the multihead attention directly draws the global dependencies of the Bi-LSTM output. Then, sentence embedding is concatenated with relation information, which comes from the KG. e_drug,e_gene and e_mutation are the drug, gene and mutation entities, respectively. v_drug−gene and v_{drug−mutation} denote the different relation vectors. Finally, sentence representation with entity relation information is fed to a softmax classifier