| Named entity recognition and relationship extraction can convert semi-structured and unstructured text data into structured data,which are used for subtasks in the field of natural language processing such as building knowledge graphs,question answering systems,and building knowledge bases.In real applications,information extraction faces many challenges.On the one hand,there is only a very small amount of annotation data for Chinese named entity recognition(NER)tasks.Chinese named entity recognition and Chinese word segmentation(CWS)tasks have many similar word boundaries.There are also specific characteristics in each task.However,the existing Chinese named entity recognition methods either do not make full use of the word boundary information in the corpus,or cannot filter the specific information in the corpus.On the other hand,for relation extraction,there is no way to make full use of all the sentences that contain information,and wrong annotations often appear.In response to these problems,this paper studies the existing Chinese named entity recognition and relationship extraction models,analyzes their advantages and disadvantages,and proposes two models.The main work content is as follows:1)Aiming at the problem of not making full use of the word boundary information in the corpus,this paper proposes a new anti-transfer learning framework to make full use of task sharing boundary information and prevent task-specific features.In addition,since arbitrary characters can provide important clues in predicting entity types,we use self-attention to explicitly capture the long-term dependence between two markers.Experimental results show that the model proposed in this paper is significantly and consistently superior to other traditional methods.Aiming at the problem of not being able to extract overlapping and multiple relations,an end-to-end sequence annotation framework based on a new decomposition strategy is proposed for the joint extraction of entities and relations.The experimental results show that the functional decomposition of the original task simplifies the learning process,obtains a better overall learning effect,and reaches a new level on the three public data sets.Further analysis shows that our model can handle the extraction of normal,overlapping and multiple relationships.3)The two models proposed in this paper are applied to the construction of power equipment diagnosis platform,and the entity recognition and relationship extraction of power equipment text are realized. |
