Research On Drug-Drug Interaction Extraction Based On Neural Networks From Texts

A drug-drug interaction(DDI)is a situation in which a drug affects the activity of another drug synergistically or antagonistically when being administered together.The information of DDIs is crucial for healthcare professionals to prevent adverse drug events.Although some known DDIs can be found in purposely-built databases such as DrugBank,most information is still buried in scientific publications.Therefore,automatically extracting DDIs from biomedical texts is sorely needed.The description of biomedical texts is characterized by professional fields,and many sentences are very long and complicated.Sometimes,in the same sentence,there are multiple drugs and also multiple DDIs between these drugs.In the face of this situation,some models based on traditional natural language processing tools and manual rules are often difficult to obtain excellent generalization ability.Although there have been some neural network based methods which can extract feature automatically for DDI extraction,the neural networks used by these models tend to be generic structures,and there is no improvement in neural network structure for DDI extraction.Main contributions in this thesis are as follows:(1)In view of the characteristics of biomedical text,a Position-aware Bidirectional Long Term Memory(P-BLSTM)model is proposed for DDI extraction.P-BLSTM can identify key information in biomedical text accurately with position-aware attention mechanism.Through the experiment on DDIExtraction 2013 dataset,P-BLSTM achieves a micro-F score of 72.25%on interaction type identification,outperforming the state-of-the-art approach by 0.78%.By adding the features of Part-of-speech,the micro-F score of P-BLSTM can be increased to73.69%.In addition,the reasons why P-BLSTM is superior to existing methods are summarized through further analysis.(2)In view of the misclassification between positive and negative instances,new models are proposed which deal with the coarse-grained classification and fine-grained classification at the same time with multi-task learning.Based on P-BLSTM,three multi-task learning networks are proposed: PM-BLSTM dual output model,PM-BLSTM dual attention model and PM-BLSTM embedding shared model.In the experiment based on the same data set and PartOf-Speech,the micro-F scores of the three PM-BLSTM were 73.69%,73.69% and 74.16%respectively,significantly higher than the P-BLSTM.We also further analyzed the differences between three PM-BLSTM models and the effectiveness of multi-task learning mechanism in DDI extraction.