Research On Relational Triple Extraction Algorithm Based On Coarse-to-fine Strategy And Confidence Factor

As big data technology gradually enters all aspects of people’s lives,there are a large number of unstructured texts in news reports,social media,blog forums and other fields,which contain a lot of valuable information.As the data continues to increase,the cost of manually obtaining valuable information is getting higher and higher,which requires machines to help us extract valuable information.Relational triple extraction aims to automatically identify entities and specific relationship types between entities.Through relational triple extraction,<subject,relation,object> triples can be extracted from unstructured text.Triples are the basis of many natural language processing applications,such as knowledge graphs,question answering systems,and more.The relational triple extraction model based on deep learning has gradually been proved in recent years that it can effectively solve the problem of relational triple extraction,but there are still some problems in the current mainstream relational triple extraction model.Most of the current models make one-time predictions for entities and relationships.Such predictions can easily confuse semantically similar relationships,thereby predicting redundant triples.In addition,the mainstream relational triple extraction data sets have the problem of category imbalance,the number of negative samples is large and easy to classify,but the current model has not optimized this.In response to the above problems,this paper proposes a new relational triple model structure and an improved loss function.In summary,the main contributions of this paper are as follows:(1)Propose a relational triple extraction model based on the Coarse-to-fine strategy:Aiming at the problem that the current relational triple extraction model predicts entities and relationships at one time,the prediction results confuse semantically similar relationships.In this paper,relational triple extraction is divided into two stages:coarse-grained labeling and fine-grained verification.In the coarse-grained labeling stage,the potential relationship between the head entity and its corresponding potential is predicted.In the fine-grained verification stage,the head entity and the potential relationship are further verified.Through two predictions with different granularity improve the prediction ability of the model for semantic similarity.(2)Proposed cross entropy based on adjustable confidence factor: The relational triple extraction model based on the Coarse-to-fine strategy uses a cascading pointer tagging scheme,which will introduce more negative samples and further aggravate the problem of unbalanced relational triple extraction tasks.Therefore,this paper starts with the loss function,and introduces an adjustable confidence factor by measuring the correctness of the prediction.In order to achieve the purpose of suppressing the loss caused by correct prediction and improving the loss caused by wrong prediction,so as to guide the model to allocate more attention to samples that are difficult to predict correctly.(3)Build a visualization platform for relational triple extraction:Based on the current wide application of relational triple extraction in the field of natural language processing,this paper encapsulates the aforementioned relational triple extraction algorithm and builds a visualization platform for relational triple extraction.The platform not only provides users with single triple extraction and batch triple extraction services,but also visually displays the extraction results to improve the efficiency of users’ access to information and facilitate the development of subsequent intelligent applications for users.