Research On Name Entity Recognition And Relationship Extraction Based On Attention Mechanism In Biomedical Text

In the field of biomedical texts,a large number of documents and cases are generated every year.These materials often contain important information,but relying on manual processing is not only a huge workload,but it is difficult to ensure efficiency.In recent years,with the continuous development of natural language processing,deep learning has replaced traditional statistical methods and has gradually become the mainstream method in various fields.Using deep learning technology to extract information from the biomedical text can effectively improve the extraction efficiency on the one hand,and on the other hand,the powerful feature learning ability of deep learning can capture more artificially difficult to identify features,improve recognition accuracy,and strengthen the extraction ability.This has a vital role in the development of the biomedical field.Information extraction can be subdivided into multiple subtasks,including named entity recognition and relationship extraction.Named entity recognition aims to mark key information in the text,while relationship extraction is to determine the connection between these entities.The main work of this paper is to solve the task of named entity recognition and relationship extraction in biomedical texts,which consists of two parts.The first part of the thesis is to design an end-to-end model of a parallel structure.Feature extraction is performed on the sequence through the parallel structure of CNN and RNN,then feature selection is performed through the attention mechanism,and finally,the labeled sequence is generated by the classifier.The model has shown superior performance in four more data sets,and two of them are even 2% higher than the current best results.The second part of this thesis is to design an feature-merger model based on the attention mechanism.The attention is used to combine the entity with the sequence,and the multiple coding sequences are merged through multiple heads.Finally,the association probability is predicted by the classifier.We also extended the sequence of entities as model inputs.Experiments have proven the feasibility of the improvement,and demonstrated optimal performance in both data sets,both exceeding the current best results by 10 percentage points.