| With the rise of big data and the improvement of natural language process capacity,text information is widely used in various scenarios.Therefore,it is necessary to understand semantic information.The process of converting semantics into a representation vector is called semantic understanding task.The current popular semantic understanding models have the following problems:poor generalization ability;poor quality of semantic understanding on small-sample datasets;the text contrastive learning lacks data augmentation methods;as the most popular application of semantic understanding,the retrieval system is difficult to balance high speed and great effect.Aiming to solve above problems,this thesis focuses on the deep learning-based semantic understanding algorithm which enhances the performance of semantic understanding effectively and designs a text retrieval system.The main contributions of this thesis are as follows:(1)Design the front projection layers and back projection layers to improve the generalization ability of the modelsIn order to improve the generalization ability of the model,this thesis designs new projection layers.The projection layers consist of front projection layers and back projection layers.Experiments verify that these projection layers can enhance the generalization of the model,obtain higher-quality semantic representation vectors,and achieve superior semantic understanding.(2)Design a graph convolution semantic understanding model to solve the small-sample problemAiming to improve the quality of semantic understanding on smallsample datasets,this thesis proposes a graph convolutional semantic understanding model that introduces graph structure with prior infonnation.It solves the problem of lacking enough information in small sample datasets.Comparing with other popular models on small sample datasets,experiments show that the graph convolution semantic understanding model can obatin semantic representation with higher quality and better understand semantic effect on small datasets.(3)Design a graph augmentation semantic understanding model to extend the data augmentationIn order to solve the lack of data augmentation methods in text contrastive learning,this thesis proposes a graph augmentation semantic understanding model including two methods of graph data augmentations,namely random edge addition and random node drop.Through experiments of different data augmentation methods,it is verified that graph data augmentation can improve the effect of semantic understanding.(4)Design a text retrieval system with balancing high speed and great effectThis thesis designs a text retrieval system based on semantic understanding model.Through the fast matching of vectors,the retrieval speed is ensured.Besides,the semantic information is obtained through the semantic representation vector,which ensures the retrieval effect.This thesis trains the model in a Chinese dataset and validates it on a practical application dataset.Experiments show that the retrieval speed and retrieval performance of the retrieval system designed in this thesis both outperform than other retrieval systems,which realize a balance between speed and effect.This thesis introduces graph model and graph data augmentation,adds the projection layers on the output of the model and improves the effect of the semantic understanding model.Besides,this thesis applies the semantic understanding model on the text retrieval system and proves the the validity of the algorithm and the rationality of its application. |
