Research On Drug-drug Interaction Extraction Algorithm Based On Deep Neural Networks

Drug-Drug Interaction Extraction(DDIE)task focuses on how to use natural language processing methods to automatically extract the interaction information between two drugs from the biomedical literature.The study of this task is of great significance for reducing accidents caused by adverse drug reactions and reducing medical costs.The main way to solve DDIE task in the past is feature and kernel function based relation extraction approach.However,this type of method often requires complicated feature engineering,which is time-consuming and laborious.With the development of deep learning neural networks,some researchers have successfully applied deep learning technology to DDIE task and achieved better results than the kernel function method.This dissertation attempts to use a variety of deep learning neural networks to solve the DDIE problem,and finally proposes a CNN-GRU model that achieves better performance than the other existing state-of-art methods in this field.Among the existing deep learning approaches for DDIE task,most of them use both the word vector and the distance vector as input,and the method usually has a simple structure and insufficient classification capability.To address these deficiencies,an improved multi-layer convolutional neural network model is proposed in the second chapter of this dissertation.The model uses only the Glo Ve word vector as input,and has multiple convolutional layers with pooling layers,which are set to improve the classification ability of the model.In this thesis,the proposed model has been trained and evaluated on the public data set provided by the DDIExtraction 2013 competition,and a F value of 74.9% was achieved,which was 3.1% higher than the current best result.The improved multi-layer CNN network model proposed in the second chapter has better classification ability,but it cannot extract the sequence information in the input text well.To solve this problem,the third chapter of this thesis improves the existing regional CNN+LSTM model and proposes a BNCNN-LSTM model for DDIE task.The advantages of the regional CNN network and the LSTM network are combined in this model,which means that,in this model,not only do the local features of each region have been sized on,but also the sequence information across regions.Therefore,the model has shown a better performance(F value 76.4%)than the improved multi-layer CNN network model in the experiment.At the same time,a batch regularization layer has been applied to the model,effectively speeding up the training of the model.The structure of BNCNN-LSTM model proposed in the third chapter is complex.In order to simplify the model,the fourth chapter of this dissertation attempts to modify the BNCNN-LSTM model,and a CNN-GRU model is proposed in this chapter.Compared to the original model,the LSTM network has been replaced with a briefer GRU network,and the batch regularization layer has been removed in the CNN-GRU model.This model has a simpler structure,as well as shows better classification performance than the BNCNN-LSTM model on the DDIExtraction 2013 dataset.