Research And Application Of Event Extraction Technology In Biomedical Field Based On Deep Learning

Event structure is crucial to capture structured biological processes in the biomedical domain and plays a vital role in bioinformatics pathway research.Event extraction is an essential subtask of information extraction that helps extract event structures from unstructured text.The event structures can be used for semantic retrieval,enriching the knowledge base,and building knowledge graphs.In this thesis,we propose a joint event extraction method based on sequence labeling for the biomedical domain.The contributions of this thesis are as follows.(1)We proposed a labeled data generation method based on pre-trained language models.Most of the existing data generation methods use distant supervision to label external corpus.However,this method may bring noise into the labeled data.This thesis addresses this problem by adopting a text-editing-based corpus generation method to generate labeled data,by applying pre-trained models to edit existing labeled data.This thesis also introduces a scoring mechanism to evaluate the quality of the generated labeled data.(2)We proposed a joint event extraction method based on sequence labeling.The pipeline method for event extraction ignores the potential connection between event triggers and event arguments.In this thesis,we propose Seq Tagee to reduce the event extraction problem into a word-level labeling problem and extract both trigger words and event arguments in a joint manner.To reduce the label space,this thesis divides labels into groups and applies a multi-task learning strategy to train the model.We use a multi-label decoder to solve the multi-label problem in event arguments extraction.Finally,the validity of the work in this thesis was demonstrated using experiments.(3)We design and implement a knowledge graph system for the biomedical domain.We obtained unstructured texts in the biomedical domain from Pub Med,and extracted event structure employing the model built in the thesis.The biomedical event knowledge graph system is built based on these event structures.The biomedical events have been visualized through this system,and the event analysis and retrieval are provided to assist researchers in their research.