Research On Drug Relationship Extraction Based On Deep Learning

As the number of Internet users continues to increase,people are increasingly demanding knowledge on the Internet.Therefore,how to collect,organize and apply this knowledge has become an important issue.The process of extracting entity relationships is the process of obtaining a lot of useful information that people need.The detailed,accurate and effective drug relationship extraction is helpful for biomedical research and development.Aiming at the problem that the traditional word vector representation method cannot fully express the relevance of related words and contains context information,a model method of fine-tuning Bert pre-training model is adopted.First of all,replace the name of the drug in the text content of the data set with many changes and expressions with the commonly used word representation.Then,through the Bert pre-trained word vector representation method,a vector representation that fully contains contextual semantic information is obtained.The resulting word vectors containing full-text semantics are flattened to obtain corresponding sentence vectors,and finally the obtained sentence vectors are added to the fully connected layer for drug relationship classification processing.Using the Bert pre-training model fine-tuning method,an F value of 70.25% was obtained when the DDIExtraction 2013 drug relationship extraction data set was used as training and test text data.In order to alleviate the problem of long-distance dependence that often occurs in text processing tasks and the problem of not paying enough attention to words that we pay more attention to.We proposes a fusion model of Bert pre-trained word vector representation,LSTM deep learning network and attention mechanism.First,use the Bert model method to obtain more accurate word vectors for semantic expression.Second,through the LSTM model that effectively solves the problem of long-distance dependence,and add an attention mechanism to focus more on key information.Finally classify the drug relationship through the fully connected layer.The experimental results show that after using the Bert + LSTM +Attention fusion model method,the f value of the drug relationship extraction data set in DDIExtraction 2013 reached 71.02%.