Design And Implementation Of Relation Extraction System In Complex Sentences Based On Few-Shot Learning

Relation extraction,which extracts structured triples from unstructured texts,is an important subtask in the field of information extraction,and plays an important role in knowledge graph construction,question answering system and other scenarios.The method of relation extraction based on deep learning relies on a large amount of high-quality labeled data,and is limited by objective reasons such as high cost of human labeling and difficulty in obtaining domain corpus.Therefore,it is of great practical significance to study the relationship extraction in a few sample scenes with limited data.Few-shot relation extraction,i.e.relation extraction in a scene with only a small amount of labeled data,has two solutions: model-based methods such as metric learning and meta-learning,and data-based methods such as active learning and so on.From the perspective of model,the research work on relation extraction with few samples simplifies relation extraction into text classification task.To be specific,it is assumed that there is only one relation triple in each data instance,and the semantic richness of the text in the actual scene is not considered.Usually,multiple entities are mentioned in each annotation instance simultaneously,and different relation triples overlap.The generalization ability of existing methods to extract relationships with few samples from the data point of view is poor,and there are some limitations because they do not take into account the situation of continuously adding relationships.This thesis puts forward a new task,Few-shot Multiple Relation Extraction,to solve the complex situation that the text contains multiple relation triples and there may be overlapping entities in the scene with few samples.Based on the setting of this task,this thesis puts forward corresponding solutions from the perspective of model and data respectively.The From One to More(FOM)model solves the problem from the perspective of model,and is dedicated to improving the utilization efficiency of model for limited labeled samples.The Actively Continual Learning-based Relation Extraction(ACL-RE)solves the problem from the data point of view,and puts forward a generalized active learning strategy to complete data accumulation.Aiming at the situation of new relationship categories in the process of continuous building of knowledge map,the continuous learning method is combined to avoid the overhead caused by retraining.The main contributions of this thesis are as follows:(1)Two datasets FSM-NYT and FSM-WebNLG are constructed for relation extraction in complex sentences with few samples,and a model for relation extraction with few samples based on target perception relation inference called FOM is proposed.The model is based on the framework of metric learning,combined with the ideas of graph convolution network and multi-head attention mechanism,which overcomes the challenges of relation inference,multi-label classification,data imbalance and NOTA(None of the above)under this setting,and improves the model.(2)The Actively Continual Learning-based Relation Extraction(ACL-RE)is proposed.This framework adopts a model of complex sentence relation extraction based on deep learning.By training neural network to simulate the expert’s decision-making process in active learning,it improves the generalization of domain on active learning and completes the data accumulation in the scene with few annotation samples.Considering the scene of new relationship categories in the process of knowledge graph construction,the model training is effectively completed by combining with the continual learning unit,so as to solve the task of relationship extraction in complex sentences in the scene of continuous few samples,and overcome the problems of relationship expansion and data accumulation in this scene.(3)Integrating the two methods and combining their respective applicable scenarios,we design and implement a complex sentence relationship extraction system based on few-shot learning to meet the different needs of users.Studying the problem of complex relation extraction in few-shot scenarios can alleviate the need for labeled data in the practical application of relation extraction scenarios,which is of great significance to the fields of information extraction and knowledge graph.