Research On Detecting Biomedical Event And Its Trigger

The twenty-first century is considered the century of biomedical technology.With the development of information technology and biomedical technology,medical literature and user-generated content carrying human knowledge and experience in natural language are showing explosive growth.It is an important research and application field in bioinformatics to obtain massive and underutilized information from various information extraction techniques.Among them,biomedical events,as n-ary relationships between entities and related concepts in biomedical texts,can more accurately and meticulously describe various levels of medical and biological functions and processes,which makes research on biomedical events a hot topic in recent years.This dissertation focuses on two basic tasks in the research of events in the field of biomedical text mining,the detection of biomedical events and event trigger words.We focus on solving the corresponding problems in the detection task,according to different characteristics of texts in social media and medical literature.For the various expressions used by users in social media texts,as well as the problems of unregistered words caused by them,this dissertation proposes a deep neural network method combined with character-level representation for event detection.This method learns the semantic,lexical,and emotional information contained in the irregular representation of social media text by character-level convolutional layers.Moreover,the multi-channel architecture,word-level representation,and other features constructed for the task are used together for the training.Experiments show that this method can alleviate the shortcomings of using word-level representations only,and it can improve the performance of event detection under different scales of pre-training word representations.For the problem of low annotated resources and high noise in social media,this dissertation proposes a biomedical event detection method based on graph embeddings.In this method,the social media text is reconstructed into a connected graph structure,so that in the process of sequence sampling on the graph,richer contextual and structural information can be learned to alleviate the shortage of annotated samples.The adversarial training method is also adopted to add small perturbations to the training samples,thereby generating a more robust detection model that adapts to the noisy input.Experiments show that this method can significantly increase the performance of event detection models based on a small-scale training corpus.Moreover,this method can bring performance improvement to different deep learning models,indicating that it has good generalization ability.For the problem of insufficient generalization ability caused by traditional biomedical event detection methods relying on feature engineering,this dissertation proposes a domain feature independent event trigger word detection method.For the problem that traditional biomedical event trigger detection methods are highly dependent on feature engineering and the lack of generalization ability caused by it,this dissertation proposes a domain featureindependent event trigger detection method.This method does not depend on the third-party feature extraction tools commonly used in previous studies or manual features designed base on the specific event type.In the feature extraction part,only a few general features such as word representation,position feature,and part-of-speech feature are used to extract higher-level feature representations using convolutional neural networks and highway networks.In the classifier part,the extreme learning machine method with stronger classification and generalization performance is used for trigger prediction.Experiments show that this method can achieve good comprehensive performance on a variety of corpus containing different types of events.For the problem of data imbalance caused a huge difference in the number of different types of samples,this dissertation proposes an event trigger detection method with controllable sensitivity.This method adopts the structure of the convolutional neural network and recurrent neural network in the feature extraction part.The convolutional neural network extracts local features and the bidirectional long short-term memory network extracts longer-span context features.In the trigger prediction part,a support vector machine method with controllable sensitivity is adopted.Through the improved objective function,the model is more concerned about the relevant features of samples from minority classes.Experiments show that this method can better balance the precision and recall of the model so that it can achieve better performance on unbalanced datasets.