Study Of Relation Extraction Methods For Multi-type Microbial Interaction

There are many kinds of interactions between microorganisms,such as symbiosis,competition,and parasitic.Studying the interaction between microorganisms is the prerequisite and basis for understanding the structure and function of microbial communities.A large number of research results on microbial interaction have been published in biomedical literature.The corresponding microbial interaction information can be extracted from these unstructured texts and organized into a structured knowledge base.This is an effective way to study microbial interactions.Existing text mining work simply defines the relationship between microorganisms as having or without,and ignores the rich and detailed types of relationships between microorganisms.Therefore,this study models the extraction of microbial relationships as a multi-type relationship extraction task,and proposes a corresponding relationship extraction model for effective microbial relationship extraction,and the results describe more abundant microbial interaction information.Specifically,the main work of this paper is as follows:(1)An integrated deep learning model is proposed for the task of multi-type microbial interaction relationship extraction.According to species interactions and semantic descriptions in the literature,four microbial interactions are summarized:positive relationship,negative relationship,association relationship,and no relationship.Then annotated the corpus(MTMICorpus)used for the extraction of multi-type microbial interaction relations,and combined with the named entity recognition tool to construct an unlabeled predictive corpus to verify the effectiveness of the model.Using this as the basis for model training,we propose an integrated deep learning model for the extraction of multi-type microbial interaction relationships.The model converts the text data into a vector representation after word vector mapping and adds position features as input.It uses three Bi-LSTM-based sub-models to independently make predictions and finally uses a voting method to make a final decision on the results of the three sub-models.Compared with traditional machine learning methods and commonly used deep learning methods,this method has achieved better results.(2)A method for multi-type microbial interaction extraction based on transfer learning is proposed.To further improve the effect of microbial relationship extraction,this paper tries to use transfer learning model to improve the effect of text modeling with the help of external domain knowledge.Specifically,we test three different pretraining models in this paper,BERT,BioBERT,and SciBERT models,which are used for the initial representation of the text,and the microbial relationship extraction task is trained through fine-tuning.The experimental results show that the model based on SciBERT has achieved ideal microbial relationship extraction results.The model is applied to large-scale unlabeled data from PubMed to predict the multi-type relationship of microorganisms.Through entity standardization processing,a total of 2855 microbial multi-type interaction relationships were obtained.This article mainly realizes the task of automatically extracting multi-type microbial interaction relationships from biomedical literature,and provides a foundation for the reconstruction of complex microbial interaction relationships.