Research On Drug-Drug Interaction Extraction From Biomedical Text

Drug-Drug Interaction Extraction(DDIE)is a task that uses text processing technology to extract the interaction information between drugs from unstructured text.The research results are of great significance to ensure the real-time update and high coverage of the drug database.This article takes the text in the field of biomedicine as the research object,and mainly performs the following related research work on DDIE tasks:(1)Aiming at the problems of inaccurate semantic information expression in the existing word embedding layer and unreasonable feature dimension setting,a word embedding model with context information(WECI)is proposed.This model uses long short-term memory networks to capture the context information of each word from the forward and reverse grammatical dependency directions,and uses a fully connected layer to fuse the word and context information to obtain each vectorized representation of the true semantics of words.The WECI model performs a DDIE task on the Sem Eval-2013 Task 9 dataset and obtains an F value of 70.14%.This performance is not only better than the benchmark word representation method 8.51%,but also higher than existing classic distributed vectors Represents the result of a method.(2)Aiming at the problems of the limited accuracy of natural language processing tools in the existing models when acquiring artificial features and the imbalance of DDIE datasets,a DDIE model based on hybrid CNN(HCNN)is proposed.The model consists of two parts: traditional convolution and dilated convolution.The former is used to capture the grammatical dependency features in continuous contexts,and the latter is used to capture the grammatical dependencies between separated contexts.This is the first time that dilated convolution have been applied to DDIE tasks.In addition,an improved objective function based on focal loss is proposed to complete parameter learning.This model achieved an F value of 75.48% on the DDIE task,which is better than the current advanced DDIE model by 1.98%.(3)Due to the limited ability of the HCNN model to divide positive and negative samples,a two-stage DDIE model based on feature fusion blocks and dynamic routing mechanisms is proposed.This method divides DDIE's original five-category task into two stages.In the first stage,a DDI detection task is performed to screen out samples with DDI.In the second stage,a DDI classification task is performed,and the sample selected in the previous step is further analyzed to determine its DDI category.The above two phases are all processed using the DDIE model based on feature fusion block and dynamic routing mechanism proposed in this chapter.This model achieved an F value of 78.08% in the DDIE task,which is superior to existing advanced and efficient DDIE models.The WECI model proposed in this paper can be well applied to various text processing tasks in the biomedical field.At the same time,the two drug-drug interaction extraction models proposed are of great significance for efficient discovery and collection of DDI information.In addition,they will help to a certain extent in the reduction of medical safety accidents and the success of drug treatment.