Research On Biomedical Dictionary-Based Entity Representation And Its Application

With the rapid development of computer technology and biotechnology,the biomedical literature is growing exponentially.Faced with these massive data,researchers are eager to reveal the biomedical knowledge contained in it,which has driven the emergence and development of biomedical text mining technology.As an important study,named entity recognition and linking aims to identify biomedical entities from text and map them to unique identifiers.As the first step of subsequent tasks such as relations extraction,information retrieval and knowledge base population,entity recognition and linking research are of great significance.There are rich dictionary resources in the field of biomedicine,which can be used as a supplement to the data-driven method to model the associated logic behind the data.Through in-depth exploration of the representation method of the entity name information,entity description information and entity structure information,this paper focuses on biomedical entity recognition and linking based on entity knowledge representation.The main content of this research are as follows:Research of entity recognition based on the combination of biomedical dictionaries and language models.Two type of dictionary features of the entity names are extracted by character matching and n-gram matching algorithms,respectively,and mapped to low-dimensional vector representations.Meanwhile,the language model is used to obtain context features.After that,we explore the impact of the combination of the above feature representations on the performance of entity recognition.Experimental results show that the dictionary feature representations help the recognition of biomedical entities,and the addition of language models can further improve the recognition performance.Research of biomedical entity linking based on entity representations learned through entity description texts.The descriptive text of the entity in the dictionary is first extracted and learned through the neural network to obtain the corresponding entity representation.The contextual features associated with the entity will then be extracted by Transformer model for entity disambiguation.Experimental results show that entity representations based on the entity description text effectively improves the entity linking performance.Research of biomedical entity linking based on entity representations learned throug entity structure information.The entity representations are learned based on an auto-encoder using the entity structure knowledge in the dictionary,including multiple variants of the same entity and different entities with the same name.The contextual features associated with the entity will then be extracted by Transformer model for entity disambiguation.Experimental results show that entity representations learned through the auto-encoder embeds the structural information between entities in the dictionary,which further improve the performance of the entity linking.This research can effectively improve the system performance for biomedical entity recognition and linking,and it is also of universal applicability and can be applied to entity recognition and linking tasks in other fields.