Research On Joint Information Extraction Methods In Low Resource Situations

With the rapid rise of deep learning and neural network technology,the field of computer vision has also entered a period of rapid development.However,in actual scenarios,traditional image recognition technology cannot meet the analysis of complex scenes.A scene graph is a medium that connects the image domain and the text domain,describing the relationship between objects and entities in the scene through a complete graph structure.However,current scene graph generation methods lack recall of high-level semantic relationships that humans are concerned about,and are difficult to meet the needs of actual scenarios.Therefore,research on unbiased scene graph generation with the optimization of object detection and predicate prediction has certain research significance and application value.This article focuses on the key technologies of scene graph generation,object detection optimization methods combined with contextual features,and long tail distribution data depolarization methods.The specific research content and contributions are as follows:Firstly,this article investigates and analyzes the relevant theoretical knowledge of scene graph generation tasks,including the basic principles of scene graph generation tasks,an introduction to object detection tasks,the principles of self knowledge distillation technology and supervised comparative learning,as well as an overview of long tail distribution problems.Secondly,this article proposes a Scene Graph Generation method with Continuous Self Study(CSS).By studying existing scene graph generation techniques,it has been found that there are two main challenges in current scene graph generation work.One is to distinguish targets with high visual similarity,and the other is to distinguish relationships with long tail bias.This article proposes a continuous self-study method(CSS),which uses a memo structure to record the behavior information of the model itself,learns its own behavior information through self-learning methods to perceive its own difficult samples,and optimizes the target based on scene context information.The experiment confirmed the effectiveness of the method.Compared with the baseline model,this method has better performance in the field of scene graph generation.Finally,in order to further optimize the differentiation with long tail bias relationships,this paper proposes a method for generating unbiased scene maps based on spatial augmentation.This method uses a fine-grained relative position encoding method to extract relative position information,and uses a supervised comparative learning method to improve the feature space of the relationship detection model.The experimental results have demonstrated the superiority of this method’s performance and the spatial enhancement ability of the relative position encoder,verifying the effectiveness of this scheme.Overall,the unbiased scene graph generation scheme proposed in this article is competitive compared to the baseline model,and has certain theoretical research significance and application prospects.