Drug-drug Interaction Extraction Based On Deep Learning

When a patient takes two or more drugs within a certain time,the efficacy of one drug may be influenced by the other.This phenomenon is called drug-drug interaction(DDI).Keeping up to date with the latest DDI information will help ensure drug safety and advance biomedical technology.The DDI extraction(DDIE)task can be divided into two parts: the natural language representation model and the semantic understanding classification model.In this thesis,a variety of deep neural network models are proposed for DDIE task.The main research contents include:(1)Aiming at solving the problems the current text representation model cannot fully express the semantics of the texts,this thesis proposes a hierarchical GRU syntax representation model.The model uses the word vectors as input,the GRU and the self-attention to obtain the semantic information of the given texts.Our model improves the performance of the natural language representation model.The model is trained and evaluated on three different datasets,which obtains 80.45%,64.21%,and 84.60% accuracy respectively.(2)Aiming at solving the problems that the existing semantic classification models of DDIE task are simple and cannot learn effective features,a deep convolutional neural network model(DDNet)is proposed to convolve the words directly and fully learn the hierarchical features in the DDI texts,which riches the DDI feature extraction process.The experimental result of F-score on the DDIExtraction 2013 corpus is 84.50%.Compared with the simple shallow neural networks,DDNet has better semantic extraction and classification ability.(3)Aiming at solving the problems that the CNN model cannot effectively learn the time sequence and the backward dependency information in biomedical texts,this thesis proposes a hierarchical bidirectional CNN-LSTM model(Bi CLSTM).The model uses two channels to process the forward and reverse texts of biomedical text separately,and then integrates the learned forward and reverse features to represent the complete semantics of the entire sentences.The experimental results show that the Bi CLSTM model achieves better performance(F-score is 87.07%)compared to the existing models in the DDIE task.This thesis proposes a variety of deep learning algorithms for natural language representation and semantic classification tasks.The rapid and accurate acquisition of DDI information from biomedical literature is of great guiding significance for pharmaceutical institutions and patients to scientifically use pharmaceuticals and drugs.