Research On Graph Codec Algorithm For Text Generation Of Triplet From Knowledge Graphs

Triple-to-text generation aims to map knowledge graph triples to text describing its information.Knowledge graphs are commonly used as information data storage in the big data era,triple-to-text generation technology can not only help to improve the productivity and efficiency of Internet data analysis,and is of great practical significance to the research and development of structured and unstructured data conversion in the field of natural language generation.Generating text from knowledge graphs is a basic task,and the completion of the task depends on the quality of the generated sentences and the accuracy of the information description.Most of the previous studies adopted standard sequence-to-sequence methods,which inevitably failed to capture the graph structure information.Although some research scholars have formulated some methodological rules for this problem,they do not have good domain generality.At the same time,how to improve the accuracy of encoding and decoding in neural networks to improve the fidelity and fluency of generated sentences remains a key problem to be solved in this task.To address these problems,two improved algorithms are proposed based on graph convolutional neural network and the relevant algorithms in the field of natural generation,and extensive experiments are carried out on the self-constructed dataset.The main research contents of this thesis are summarized as follows:1)Aiming at the problem of how to improve the model encoding graph information,this thesis studies how to utilize the convolutional neural network as the encoder of the model to replace the general linear encoder for preserving the structural information of the input graph itself,and proposes a graph enhancement planning algorithm(GEP).The algorithm embeds each relation and word in the input triplet as a new node to enhance graph information,and reduces the difficulty of model embedding and feature extraction.2)Based on the research of graph convolutional neural network encoder,a feature fusion algorithm is proposed,which employs global node encoding and local node encoding in a serial manner to retain the information and advantages of both encoding computation methods.At the same time,this paper studies the multi-layer stacking of graph convolutional neural networks to obtain deeper feature expressions,and ensures the gradient stability and hidden state information transfer of the model during the training process by adding dense connections.3)Aiming at the problem of how to obtain higher quality generated tex t,this thesis makes various algorithm integration improvements to the decoder,and integrates the context gate into the LSTM network to preserve the context information during the hidden state update process,thus ensuring faithful original meaning.Meanw hile,this thesis add replication attention and penalty mechanisms in the decoder to address out-of-vocabulary(OOV)problems and improve the quality of the generated sentences.In this thesis,the performance of the model is verified on the English data set Web NLG and the self-built Chinese marine industry dataset.Extensive experiments show that the proposed methods can effectively generate high-quality text from the graph structure input,and achieve high scores in the four automatic indicators.Further more,the effectiveness of the proposed algorithm is also demonstrated in the ablation experiments,which can effectively improve the performance of triple-to-text generation task.