Research On Joint Extracting Models Of Named Entities And Relations Based On Deep Learning

Named entity recognition and relation extraction are two main information extraction tasks in the field of natural language processing.They are designed to extract entities and relations from unstructured text.The relation triples composed of them can be directly applied to the construction of knowledge bases.At the same time,it also plays a vital role in upper-level applications such as information retrieval and question answering systems.After in-depth analysis of the design ideas and modeling mechanisms of the existing algorithms,it was found that these methods have the following shortcomings: First,the traditional methods process named entity recognition and relation extraction in a pipeline in two steps.Two sets of different model structures and parameters are used to split the internal connection between the two tasks.When the upstream task fails,it will inevitably lead to the downstream task error.Second,because there may be multiple relation triples in a piece of text,entities may overlap between the relation triples.At the same time,the relation between a pair of entities is also directional.Most of the existing modeling methods cannot take into account these two challenging problems at the same time.This thesis focuses on the joint extraction model of entities and relations,considering the overlap of entities and the directionality of relations,and proposes corresponding solutions.The main work content and contributions are as follows:(1)A specific relation representation method is proposed.This method can help the target output to convert between the set of relation triples and the text sequence in both directions.The encoder-decoder model with attention mechanism is successfully applied to the entity and relation joint extraction task.Entity recognition and relation extraction are integrated into the decoding stage to freely generate entity and relation information.The joint extraction method not only fully captures the interaction between entities and relations,but also effectively solves the problem of entity overlap,and clearly specifies the direction of the relationship between the entity pairs.Compared with existing similar model methods,it has more advantages in the completeness of entity mention.Through the performance evaluation on the public dataset,the experimental results prove the effectiveness of our method.(2)A pre-trained language model-based entity and relation joint extraction framework is proposed.The framework consists of two parts: entity recognition module and relation detection module.They establish task associations by sharing the same pre-trained language model,and jointly train the two modules through cross-sampling to find the best model parameter balance globally.In the relation detection module,each pair of entities and their order are considered,and the problems of entity overlap and relation directionality are solved at a more granular level.The experimental results on the public dataset show that the joint extraction method is superior to the existing research work,which proves the performance superiority of our method.(3)We apply the joint extracting models of entities and relations to the alarm text,propose a set of feasible solutions to extract the entities and relations of the alarm text,formulate data labeling standards and business logic processing procedures,and finally complete the development of the corresponding prototype system.We hope to promote the practical application of this method in related text processing fields.