Research On Biomedical Entity Relation Extraction

Biomedical entity relation is the cornerstone for the acquisition of various kinds of biomedical knowledge,and also an essential part of complex biochemical network.With the development of biomedical technology and information technology,research on biomedical entity relation extraction has increasingly become one of the most popular hotspots in many interdisciplinary fields.Therefore,it is of great significance to rapidly and reliably extract entity relations of interest from biomedical free text.This research mainly focuses on the document-level relation extraction between chemicals and diseases from biomedical literature in order to better serve the research,practice and industry in biomedical field.The main contributions are as follows:1.Firstly,we proposed a feature-based machine learning method with hypernym-hyponym filtering and multi-level linguistic features to extract biomedical entity relations at document level.According to the co-occurrence of the entities participating in the relations,we factorized the relations at document level into two different mention levels,i.e.intra-sentence level and inter-sentence level.We then employed various linguistic features for the relation extraction at both levels,respectively.Finally,the relations from different levels were fused to acquire ultimate relations at document level.During the relation extraction stage,we leveraged a hypernym-hyponym filtering method to obtain more specific relations in order to ensure the consistency among multiple relations.The experimental results show that the feature-based method with hypernym-hyponym filtering and multi-level features is effective for document-level relation extraction in biomedical literature.2.Secondly,we proposed a machine learning method based on the representation of contextual and dependency information for document-level biomedical relation extraction with the emphasis on intra-sentence level relations.We improved the feature-based method by representing the contextual and dependency features with a convolutional neural network for relation extraction at intra-sentence level.Deep learning was aimed to represent the relation instances in a more abstract way and capture the deeper semantics for entity relations.Experimental results show that the machine learning method based on the convolutional neural network with contextual and dependency information can significantly improve the performance of the document-level relation extraction from biomedical literature.3.Finally,we proposed a machine learning method based on distant supervision to extract biomedical entity relations at document level in order to alleviate the issue of insufficient training data.We first leveraged enormous biomedical knowledge from databases to automatically construct training relation instances from biomedical literature.We then proposed an attention-based recurrent neural network and a stacked autoencoder neural network for the relation extraction at intra-and inter-sentence level,respectively.Experimental results show that the machine learning method based on distant supervision is able to reach the state-of-the-art performance of the document-level relation extraction from biomedical literature.In conclusion,this work is devoted to document-level biomedical relation extraction.On one hand,we proposed effective methods to improve the performance of the biomedical relation extraction.On the other hand,we also tried to promote the research progress in biomedical information extraction.This research has achieved some preliminary results,which we hope can not only be helpful to other researchers but also promote the development of the deep natural language understanding in biomedical area.