Research On End-to-end Joint Knowledge Extraction Based On Sequence Labeling

Named entity recognition and relationship extraction is an important part of knowledge extraction,and it is also a key technique for building a knowledge graph.At present,there are two main research methods.One is to use the pipeline method for extraction,first performing entity recognition and then performing relationship extraction tasks.However,the former error will affect the performance of the following relationship classification,and will also produce a lot of unrelated entities for this redundant information.The second is to use joint extraction,while extracting entities and their relationships from unstructured text,which can make up for the former.However,currently it is also a feature-based structured system that requires complex feature engineering,which may also lead to error propagation.This article has conducted in-depth research around entity recognition and joint joint extraction,the specific contents are as follows:(1)The named entity recognition and relationship extraction tasks were transformed into sequence labeling tasks,and a new end-to-end joint extraction model based on a special labeling strategy was proposed.This model uses Google's open source BERT model to train distributed word vectors as an initialization method for label semantic representation to represent input text.It then encodes with a bidirectional LSTM layer,and decodes with an LSTM layer that incorporates the attention mechanism.Finally,the predicted label results are extracted through the output of the softmax function.In addition,the model uses an objective function with a bias to optimize the model so that the model can predict the entity-to-relationship more accurately.On the NYT public data set,the model has achieved relatively good performance.(2)The knowledge graph technology is applied to the related transaction announcement,and an attempt is made to establish a lightweight knowledge graph for this field,which is helpful for the enterprise to analyze the transaction.Related party transaction announcements,as one of the important types of announcements for listed company information disclosure,have strong research value.This article explains the processes of ontology construction,knowledge extraction,knowledge storage,and visual analysis in the construction of the knowledge graph of connected transactions.In summary,the innovation of this paper is to propose a new end-to-end joint extraction model based on labeling strategy for entity recognition and relationship extraction tasks.