Extracting Biomedical Entity Relations Via Graph Neural Network

As a frontier interdisciplinary discipline,biomedical science is closely related to human life and health.In recent years,with the rapid development of the biomedical field,the amount of biomedical literature has been rising exponentially.The rich information hidden in these data is significant with regard of the biomedical field of drug discovery,disease prevention and knowledge graph construction.Therefore,processing and analysing unstructured biomedical literature via text mining could significantly accelerate the research development in this field.As an important branch of text mining technology,relation extraction can automatically extract information from unstructured text.At present,the mainstream relation extraction method is based on deep learning.However,it is difficult to effectively learn the semantic and syntactic information of the complicated sentences in biomedical literature for some neural networks.By contrast,the graph neural network can directly operate on the graph and obtain the representation that contains the information of each node and neighbor nodes in the graph.Hence,we firstly propose a graph neural network based method for the chemical-protein relation extraction.This method can effectively learn the sequence information and longdistance syntactic relation of sentences based on the dependency parse graph.The experimental results indicate that the method achieves competitive performance in chemical-protein relation extraction.Due to the long and complicated feature of biomedical texts,there are always many noisy words in the sentences,which could impact on the result of relation extraction model.To alleviate the problem,we propose a multi-head attention based graph neural network model in this paper.The multi-head attention mechanism can automatically learn the correlation between each node in a sentence according to contextual information.The full-connected weighted graph based on the correlation between nodes not only preserves the complete information of the sentence,but also enables the model to focus on the relevant substructures for the relation extraction task.Therefore,the graph neural network can effectively reduce the interference of irrelevant information when learning the key information based on the full-connected weighted graph.The experimental results in multiple biomedical relation extraction tasks show that our method can effectively reduce the impact of noisy data and further improve the performance of relation extraction system.In biomedical relation extraction,some relation entity pairs exist in different sentences.Compared with intra-sentence relation extraction,cross-sentence relation extraction needs to consider a wider text range.Therefore,it is more difficult to extract.For cross-sentence relation extraction,we propose a method based on integrated graph neural network.By using two parallel graph neural networks,the method can simultaneously learn on dependency parse graph and full-connected graph to capture more semantic and syntactic information within and between sentences.In addition,in order to explore the application of language pre-training model in biomedical entity relation extraction,we obtain the words vector representation from language pre-training model Bert,and the experimental results show that the proposed method achieves satisfactory performance in cross-sentence relation extraction.