Research On Join Extraction Of Resource Entities And Relations Based On Deep Learning

Resource named entity recognition and entity relationship extraction are the important basis for semantic information understanding of resource description texts.Based on entities and relationships,resource database and resource knowledge map can be constructed,which is of great significance for the in-depth study and full utilization of resources.At present,entity recognition and relation extraction usually adopt statistic-based model or artificial neural network-based model,and most studies regard these two processes as separate models,The two processes are completed sequentially.Because the two modules have similar underlying data processing process,this method is easy to cause repeated data preprocessing,but also has the problem of error propagation.The wrong entities in the entity recognition phase will continue to be transmitted to the relationship extraction module,affecting the determination of entity relationship.In view of the current situation of the problem,the following work and innovations have been done:(1)Overcoming the excessive dependence on artificial features,this paper constructs a resource entity and relationship recognition model based on depth learning and rule combination,extracts context features through Bi-LSTM-based circular neural network,and then uses CRF to complete entity recognition and CNN model to complete relationship extraction;Because the rule matching method has a high accuracy rate,this paper complements the advantages of the depth learning model and the rule model,and supplements and verifies the recognition results of the depth learning model by formulating a small number of rules,so as to improve the accuracy rate.(2)A hybrid neural network model with mutual feedback mechanism is proposed to extract entities and their relationships jointly,Firstly,the two modules share the same process of word vector transformation and context feature extraction to avoid repetitive data preprocessing;secondly,the feedback check of entity recognition by entity relation extraction is implemented to enhance the relevance of the two models,restrain the error propagation in the two stages and improve the overall recognition performance;thirdly,the mutual feedback mechanism is used to enhance the correlation between the two models.(3)Thestructure of resource knowledge representation is defined,and an algorithm for constructing resource knowledge set is proposed.The algorithm generates resource knowledge set which is easy to store and manage by extracting entities and relationships.The proposed algorithm is applied to entity dataset for experiment,Under the same hardware and software environment,the proposed method can shorten the model training time,improve the precision rate,recall rate and F1 value of entity and relation extraction,increase the F1 value of joint extraction by 3.91%,increase the F1 value of entity recognition sub-module by 1.34%,increase the precision rate of relation extraction by75.02%,increase the F1 value by 5.79%,and obviously improve the effect of entity relation extraction.The experimental results show that the joint extraction model can achieve the combination of the two sub-modules to reduce data processing time and error data transmission,and the mutual feedback mechanism can improve the overall recognition effect.