Chinese Text-oriented Medical Knowledge Acquisition,Representation And Reasoning

Text-oriented knowledge acquisition,representation,and reasoning lies in the cross field of natural language processing(NLP),knowledge representation,and knowledge reasoning(KRR).This topic mainly focuses on the problems such as automatically extracting triples from the text,manipulating the triple information into knowledge,and reasoning new knowledge from the existing ones.These problems aim to imitate the indispensable abilities of agents to acquire information from the environment,generate corresponding knowledge,and inferring the unknown from the known,which are the cores of artificial intelligence.With the development of recent 20 years,text-oriented knowledge acquisition,representation,and reasoning has started to see some success in the open domain,which is mainly reflected on the development of several of the knowledge extraction methods(e.g.,bootstrapping,distant supervision),the release of a number of large-scale knowledge bases(e.g.,YAGO,NELL),and the rise of some tractable knowledge reasoning methods(e.g.,path-ranking,Trans X).In recent years,the rise of deep-learning-based methods opened new opportunities for this field,and it has been demonstrated in improving the performance of existing systems.For the specific domains,the general methods of open domain usually combine with the domain specialty,to benefit from the depth of the domain knowledge,as well as avoiding several difficulties in the open domain(e.g.,entity disambiguation).Knowledge in specific domains plays important role in both professional areas(e.g.,medicine,law,biology)and living areas(e.g.,entertainment,food,business).This work focuses on the medical domain,and study the key techniques in knowledge acquisition,representation,and reasoning for Chinese electronic medical records,and the semi-structured medical text.This work consists of following three parts:(1)Medical knowledge acquisition and representation from CEMR.To address the poor performance of medical relation extraction systems,we propose medical knowledge network(MKN),a network only contains medical entities but not the entity relationships,to represent the medical knowledge in Chinese electronic medical records(CEMR).Nodes of this network are medical entities and edges are co-occurrence relationships between entities in the same records.Although the co-occurrence can hardly be regarded as the actual medical knowledge,MKN indeed shows the knowledge complexity compared with the random networks,which indicates the possibility of building the MKN-oriented knowledge reasoning systems.(2)Medical knowledge acquisition and representation from semi-structured medical texts.As a single knowledge source,EMR can hardly cover all the medical knowledge needed in clinical practice.To extract the medical knowledge from other sources,we propose an architecture to automatically extract triples from semi-structured medical texts,and then integrate the triples as a medical knowledge graph.It contains the knowledge acquisition module(medical entity recognition,and entity relation extraction)and the knowledge manipulation module(knowledge description,storage,normalization,and verification).(3)MRFs-based clinical decision support.Focusing on the fact that MKN is organized using the co-occurrence associations,we propose a medical knowledge inference method based on the Markov random fields(MRFs).Focusing on three clinical decision support problems(diagnosis,test recommendation,and treatment recommendation),we design six energy functions based on the graph-based features,edge-weight features,and distributed-represented node features.Three evaluation measures are adopted or designed to evaluate the performance of the system.Experiments on actual EMR data demonstrated the efficiency of our method to select the knowledge from MKN.Knowledge graph can promote this process by entity normalization and association explanation.In conclusion,focusing on two kinds of medical texts,we study the corresponding key techniques regarding the knowledge acquisition,representation and reasoning.The effectiveness of proposed methods has been experimentally demonstrated on actual EMR data and medical texts.The research results are preliminary but still meaningful.We hope this work can be applied to more varied medical text and broader reasoning tasks,and further promote the research and development of NLP and KRR in the medical domain.