Research On Biomedical Event Extraction Technology

Biomedical event extraction technology can help researchers quickly locate events accurately from a large number of biomedical literatures and express them in a structured form.It has important research significance and application value in drug development,clinically-assisted diagnosis and treatment,and construction of biomedical ontology database.Biomedical events describe the process of biomedical entity state changes,mainly composed of triggers and event arguments.This paper focuses on the key technologies of biomedical event extraction,and focuses on the techniques of trigger recognition and event argument recognition based on deep learning methods.The main research contents of this paper are as follows:(1)Combined self-attention mechanism for biomedical event trigger recognition.Triggers represent the actions(inhibition,development,formation,etc.)and the type of event.Aiming at the problem that there are a lot of complex events in the text of trigger recognition task,which leads to insufficient feature mining in the text,this paper proposes a method of event trigger recognition that incorporates self-attention mechanism.The method focuses on trigger recognition in text articles,the introduction of the words of the distributed based on the semantic features,by using the Bi-GRU to capture relevant context features,moreover,self-attention mechanism is integrated into the Bi-GRU model to fully mine the features contained in the text from multiple levels,so as to better understand and express the semantic information expressed by the sentence.Finally,the word-level features are fused to classify the triggers through the classification layer.(2)Biomedical event argument recognition based on C-BiGRU and attention mechanism.Event arguments are the participating arguments involved when an event occurs,mainly biomedical entities or other events.In this paper,the task of event argument recognition is regarded as a relational extraction task,which identifies and classifies event arguments and their types.In the event argument recognition task,a sentence may have multiple triggers,and a trigger may correspond to multiple arguments,so it is necessary to extract deeper semantic features for recognition and classification.This paper proposes an event argument recognition method based on joint neural network,which introduces dependency information and adds domain word embedding,distance features,event and entity type features to enrich semantic features,combining CNN and Bi-GRU models for effective extraction of deep features,and attention mechanisms are introduced to focus on more important feature information and improve classification accuracy.Finally,according to the definition of events in the biomedical field,the recognition results of event arguments are processed with rules to generate the final structured events.Through experiments on MLEE data set,it is found that,compared with existing methods,the proposed method in this paper achieves higher comprehensive performance in both task of event trigger recognition and event argument recognition,which proves the rationality and effectiveness of the proposed method in these two tasks.