The Research Of Methods For Microbial Interaction Extraction Based On Medical Literature

Microorganisms inhabit in various ecosystems on earth,and they actively interact with community members to maintain environmental balance and stability.The construction of microbial interaction network is one of the most effective methods to characterize the interaction between microbial community members and the host.In particular,the microbial interaction extraction(MIE)is the basic work and key steps of microbial network construction.With the publication of a large number of medical literatures,many experimentally verified microbial interaction information is dispersed therein.It is important to extract this information and organize it into a database or knowledge map.Text mining technology is able to automatically extract and integrate these microbial interactions in medical literature.In this paper,we focus on the task of Microbial Interaction Extraction(MIE)on texting mining,and the main contributions are as follows:.Firstly,a supervised learning method based on Support Vector Machine and feature vectors is proposed for MIE.In order to train the model,we manually labeled a microbial Interaction corpus,containing 1005 abstracts and 7483 entity pairs,which provided an data resource for MIE.We designs 41 kinds of features,which can be divided into four categories:word features,part of speech features,dependency features and logical features.The optimal feature subset is obtained by feature selection,and the feature vector is constructed as the input of Support Vector Machine model.Finally,the effects of various characteristics on the performance of the MIE system are discussed and analyzed.The experimental results show that the optimal feature subset including lexical feature,dependency feature and logical feature,which lead to a effective MIE system.Secondly,a semi-supervised learning method based on word representation and neural network is provided for MIE.The design of performance of relational extraction system based on supervised learning method requires domain expertise and linguistic background,which has a good universality.The word representation of deep learning provides can automatically summarize effective characteristics from a large amount of data.This paper proposes a method based on word representation and neural network,which can automatically find the feature information in the text for effective data representation,thus simplifying the complex feature engineering in traditional machine learning methods.On the basis of Word2vec to train the word vector resources related to microorganism,the Max-Bi-LSTM model is constructed aiming at the problem of extraction of microbial interaction.Several experiments are designed and other commonly used models in biomedical text mining are compared and analyzed.Experimental results show that the proposed model has good performance.In the end,a system that can automatically extract microbial interaction from medical literature is constructed by combining the named entity recognition model and the relationship extraction model.Our research provides a method and tool for constructing an efficient microbial interaction extraction system.