Research On Relation Extraction Based On Contrastive Learning

With the rapid development of deep learning,relation extraction has made great progress,but its performance depends on the quality and quantity of annotated data,causing a certain degree of trouble to the actual application.Zero-shot relation extraction aims to identify unseen relations by training on seen relations,which can alleviate the dependence of model on annotated data to a certain extent.Although zero-shot relation extraction can deal with unseen relation,it cannot deal with relations without predefined.For non-predefined relations,open relation extraction is needed to solve the problem.For unseen relations and non-predefined relations,this paper conducts in-depth exploration and research on zero-shot and open relation extraction scenarios under relation extraction based on contrastive learning.Specific research contents and contributions are as follows:First,this paper investigates and sorts out the relevant theoretical knowledge and research status of zero-shot and open relation extraction,and introduces the principle of contrastive learning,data augmentation and loss functions.Secondly,in view of the similarity problem that existing methods of zero-shot relation extraction are limited by,this paper proposes a new relation contrastive learning framework RCL based on contrastive learning.Under the setting of multi-task learning,this paper introduces relation contrastive learning,and trains on seen relations through relation contrastive learning and relation representation learning,so that the model can learn the characteristics of the relation itself,the difference between relations,and the difference between instances,so as to generalize this ability to the extraction of unseen relations.Experimental results show that RCL can effectively solve similar problems.Thirdly,Aiming at the problem that the existing methods of open relation extraction are limited by the pseudo-correlation of datasets,this paper proposes an open relation extraction model based on contrastive learning.Under the setting of multi-task learning,the model can model the relative relation constraint of positive,hard-negative and semi-hard negative instances by instance ranking and label calibration,so as to learn a more discriminative relation representation and improve the performance of downstream clustering.Experimental results show that the model proposed in this paper exceeds all the baseline models compared,which proves the effectiveness of the method.Finally,this paper summarizes the research content,and looks forward to the future development of zero-shot and open relation extraction,and gives several future research directions.