| In the field of biomedicine,a large amount of biomedical literature is published daily,and the literature contain many new interactions between drugs,diseases,and symptoms,which is an important resource for biomedical research.As of February 2021,Medline,an international comprehensive bibliographic database of biomedical information established by the U.S.National Library,has more than 27 million titles.The efficient and accurate extraction of structured biomedical knowledge from this massive biomedical literature is of great significance to accelerate research in biomedical-related fields.Currently,the research on relationship extraction in biomedical fields mainly relies on supervised learning methods,which are highly dependent on annotated data.However,building high-quality biomedical annotation data is not only time-consuming and laborious,but also requires annotators with certain biomedical domain knowledge.Therefore,the lack of annotated data in biomedical domain greatly limits the practical application value of biomedical relationship extraction research.In this thesis,we introduce adversarial domain training and biomedical domain knowledge sources to reduce the dependence of biomedical relationship extraction models on labeled data and improve the relationship extraction performance of the models with few labeled samples.(1)Biomedical relationship extraction based on adversarial domain adaptationFor the situation that the target task labeled data(target domain data)is relativ ely small,the extraction performance of the model on the target task is improved by introducing the adversarial domain adaptation learning method to utilize the similar task dataset(source domain data)with sufficient labeled data.Due to the distribution differences between target domain data and source domain data,this thesis adjusts the parameters of the model based on the adversarial domain adaptation strategy to adapt to the distribution characteristics of the target domain data,improves the performance of the target domain task using the source domain labeled data,and reduces the dependence of the model on the target domain labeled data.(2)Biomedical relationship extraction based on knowledge modelThere are relatively complete domain knowledge sources built in the biomedical field,and the structured knowledge in these knowledge bases is highly indicative for biomedical relationship extraction in related fields.In this thesis,based on the knowledge representation learning method,we learn the knowledge representation of the triples in the biomedical knowledge base to obtain the knowledge representation of entities and relationships;we use the splicing method to fuse the learned knowledge information into the biomedical relationship extraction model,and reduce the dependence of the biomedical relationship extraction model on the labeled data by introducing the relevant domain knowledge.The biomedical relationship extraction method based on adversarial domain adaptation and the biomedical relationship extraction method based on knowledge model proposed in this thesis are validated and evaluated on international publicly available biomedical relationship extraction datasets,respectively.The experimental results show that the method proposed in this thesis can effectively reduce the dependence of the biomedical relationship extraction model on labeled data by introducing relevant task labeled data and biomedical domain knowledge. |
