Research Of Few-shot Relational Triple Extraction Model

With the rapid development of the Internet generating a huge amount of data,how to extract valuable information from it has become an urgent problem,and information extraction research has developed rapidly.Entity relation extraction is a hot research and application technique to extract information of relational triple(subject,relation,object)from unstructured text and construct knowledge graph.Mainstream entity relation extraction methods rely on a large number of annotated training samples,however,practical applications often lack the existence of cold-start problems.How to use a small number of annotated samples for relation triple extraction has become one of the hot topics in this field.Most of the existing works focus on the recognition of subject-object relations,and the joint extraction of triples with few-shot is in the initial stage,and the accuracy of these methods for subject-object recognition is low.To address the problems,this paper focuses on the joint extraction of triple in the case of few-shot.Firstly,we propose a template-based nearest neighbor matching few-shot entity relation extraction model TNM for the triple extraction in sentences.Secondly,an improved entity relation extraction model NNM with few-shot is proposed to further improve the accuracy of subject-object recognition.Finally,the NNM model is applied to multi-domain datasets,and the transfer learning mode is adopted to fine-tune the NNM model.The main contributions of this paper are as follows.(1)Combining the named entity recognition method with the relation classification method,a template-based nearest neighbor matching few-shot entity relation extraction model TNM is proposed to achieve end-to-end triple joint extraction.For subject-object recognition,a set of subject-object candidate templates are constructed manually,and the similarity between the candidate words and the templates is calculated for prediction;for relation classification,the prototype representation of each relation is calculated,and the relation prediction of sentencecontaining triple is performed based on the nearest neighbor idea.Experiments on the publicly available dataset Few Rel demonstrate the effectiveness of the method.(2)In order to avoid the construction of manual templates,this paper further proposes an improved few-shot entity relation extraction model NNM,which automatically combines the subject-object label distribution pattern of data in the source domain to build a classification model when identifying the subject-object.the NNM model identifies the subject-object in a sentence based on the similarity between words and the distribution of labels between the subject-object,and uses the similarity of subject-object entity types as loss to improve the recognition accuracy.Experiments show that this method substantially improves the accuracy of subject-object recognition and achieves the best results so far in terms of the accuracy of triple group recognition.(3)In this paper,the NNM model is directly applied to the Chinese and English datasets from other fields,and achieves similar results as the training dataset,which indicates that the method has good generalization.In addition,this paper adopts the transfer learning model of "pre-training with unlabeled samples-pre-training with labeled samples-fine-tuning in NNM model",and fine-tunes two source models with different network structures and number of parameters trained in the publicly available large sample English corpus on the NNM model respectively.The experiments show that the NNM model works better when the source model with a small number of parameters and a simple network structure is chosen for new domain applications.