Drug-drug Interaction Extraction Model Based On Bidirectional Long Short-term Memory

Drug-drug interaction extraction is an application of entity relation extraction in the biomedical field.It is very important to domains such as biomedical text mining,drug side effects alert and drug information retrieval.Existing models mostly only utilize semantical information or syntactical information from the text itself.They often ignore the biomedical domain-specific knowledge or introduce out-of-date biomedical information.During the integration of representations,existing methods just concatenate representations together and omit relations among them.In order to solve the absence of the biomedical domain-specific knowledge,this thesis proposes a drug-drug interaction extraction model which introduces external biomedical resource.In the phase of text representation conversion,besides word embeddings and entity offset embeddings,the model adds a type of low-dimensional distributed external biomedical resource representation: concept embeddings.Concept embeddings are trained from the external professional corpora and contain semantical information of biomedical entities.Lastly,converted text representation sequences are fed into a bidirectional long short-term memory based deep learning classifier to get final predictions.In order to solve the weakness of out-of-date information and the omission of relations among representations,this thesis proposes a new drug-drug interaction extraction model which inserts user-generated content embeddings and full attention representation merging mechanism.In the step of representation generation,the model adds a kinds of user-generated content embeddings to act as timely biomedical representations.At last,text representation sequences are fed into a bi-directional long short-term memory and Transformer based deep learning classifier to output predicted labels.In summary,this thesis aims to solve the absence of biomedical information and the issue of out-of-date biomedical knowledge.It introduces two kinds of external biomedical resource representations: concept embeddings and user-generated content embeddings.In order to solve the omission of relations among representations,this thesis proposes a kind of attentive representation merging method.Models proposed in this thesis improve the performance of drug-drug interaction task and overcome the weakness of existing methods.