Biomedical Event Extraction For Nested Structure

With the continuous development of biomedical research,the number of biomedical literature has grown tremendously.How to automatically extract useful information for researchers from a large number of biomedical literature and present them in a structured form is a challenging task to be performed by information extraction in the biomedical field.These structured information,such as relations and events between biomedical entities,etc.,has important application value for the construction of biomedical knowledge graphs and new drug development.In the biomedical field,an event consists of a trigger and several arguments,where the arguments can not only be biomedical entities but also other biomedical events,and in this case the event is in the nested structure.The main research topics in this paper are as follows:First,the pipelined biomedical event extraction based on n-ary relation extraction.For the uncertainty of the number of arguments in an event,this paper proposes to adopt an n-ary relation extraction method to determine the correctness of candidate events with multiple arguments in order to capture the semantic relationship between a trigger and its arguments,and to achieve the assembly of such events.From the experimental results,it can be seen that the method can significantly improve the performance of biomedical event extraction.Second,the pipelined biomedical event extraction based on prior knowledge.In order to better utilize the knowledge in the biomedical domain,this paper adopts an a priori knowledge-based approach for the pipelined biomedical event extraction,where the related information in the corpus annotation guideline documents is transformed into various prior knowledge to assist each sub-task.The experimental results show that appropriate prior knowledge can improve the performance of each subtask,which leads to the improvement of the overall performance of event extraction.Third,joint extraction of biomedical events based on multi-layer sequence labeling.In order to alleviate the problem of error propagation in the pipelined biomedical event extraction,this paper proposes a multi-layer sequence labeling approach for joint learning of trigger word recognition and argument role recognition,so as to address the nested structure in biomedical events and to capture the interrelationship between subtasks.Experiments show that the joint extraction method based on multi-layer sequence labeling can achieve better performance and also improve extraction efficiency.