Biomedical Information Extraction Based On Event Framework

With the popularization of electronic documents and the improvement of computer performance,it has become commonplace to use machines to extract the knowledge of biological information and to construct entity relation networks.Event extraction task contains two sub tasks,event trigger identification and event arguments detection.In the field of Natural Language Processing,with the development of deep learning and representation learning,more and more tasks use the method of deep learning instead of the traditional method based on Feature Engineering,thus avoiding the extra work of artificial extraction of features and the introduction of too much prior human knowledge to limit the model.On the basis of these works,this paper combines deep learning and biomedical event extraction,and proposes deep learning models to realize trigger identification and event arguments detection.For the task of event trigger identification,there are two methods proposed.The first proposed method abstracts this problem into the task of word classification.On the basis of the distributed semantic space,a Dynamic Piecewise Max-pooling Convolutional Neural Network(DPM-CNN)model is proposed in this paper.This model can obtain the sentence level feature combined with location information for the candidate trigger word,both in lexical level and entity level.Finally,a classifier will be trained to identify biomedical event triggers.The proposed method effectively combines the position information of the candidate trigger word with the model structure.The experimental results show that the proposed method can improve the trigger identification performance compared to the traditional feature engineering method based on SVM.The second proposed method abstracts the trigger identification task into a sequence tagging task,and proposes a method based on Bi LSTM-Attention-CRF to realize trigger identification.This method annotates the samples in the form of BIO labels so the model can identify triggers formed by multiple words in the text.Then the bidirectional LSTM network is used to construct the basic feature where document level attention will add on.Finally,the condition random field is used to learn the correlation between BIO tags and then annotate the current candidate words in the sentence to extract the trigger words in the text.For the event arguments detection task,a sequence labeling method is proposed to detect event arguments in this paper.Unlike the traditional classification method,this paper abstracts the problem into a sequence labeling task for the word sequence where the trigger word is in.Because the event arguments detection is closely related to the trigger word that triggers the event,the sentence of the trigger is taken as the basic sequence information,and then combines with the entity category and the trigger word category information in the sentence.Finally,the event arguments is labeled in the sequence as the results of event arguments detection.The experiments are carried out on the MLEE data with the PubMed literature summaries to train word embedding representation,and construct the corresponding distributed feature representation for different methods.The Precise,Recall and F1 score are used to evaluate results.The experiment results verify the effectiveness of the methods proposed in this paper.