Biomedical Relation Extraction Based On Neural Network And Domain Knowledge

Information extraction technology provides means to deal with massive data on the Internet.Biomedical literature is an important part of text data on the Internet,and the knowledge hidden behind the literature can be used in the practical applications such as information retrieval,recommendation system and question answering system.Relation extraction is necessary for mining the knowledge contained in the literature.This master dissertation focuses on the document-level biomedical relation extraction task.Document-level relation extraction is different from traditional sentence-level relation extraction.Sentence-level relation extraction extracts entity relation in a single sentence,but document-level relation extraction extracts entity relation in an abstract.An abstract consists of several sentences and it has more complex semantics.This dissertation proposes an end-to-end relation extraction model based on graph convolutional neural network and multi-head attention.This dissertation constructs a document-level dependency graph and uses graph convolutional neural network to obtain syntactic structure features of text.On the other hand,the multi-head attention mechanism is employed to learn the relative important context features from different semantic subspaces.To enhance the input representation,the deep context representation is used in our model instead of traditional word embedding.This dissertation proposes a relation extraction model based on domain knowledge.In addition to textual information,existing biomedical domain knowledge can also provide more information for relation extraction.This dissertation explores how to obtain domain knowledge from existing biomedical knowledge database and how to effectively combine domain knowledge and textual information for relation classification.The biomedical knowledge database is used to construct a large number of knowledge triples,and knowledge representation model is used to learn the domain knowledge contained in the triples.Entity representation and relation representation are obtained from the knowledge representation model.The relation representation vector is used as knowledge vector,and attention mechanism is used to integrate domain knowledge with textual information to further improve the performance of relation extraction task.The research is carried out on the chemical-disease relation extraction dataset proposed by BioCreative-V.Sufficient comparative experiments are designed and completed to verify the effectiveness of the model proposed in this dissertation.The experimental results show that the proposed model can effectively improve the performance of the relation extraction task.