Research On Text Classification Of Rural Tourism Evaluation Based On Graph Convolution Neural Network And ERNIE-gram

Nowadays,more and more people tend to travel to popular attractions,and before each trip,they tend to take the evaluation of the attractions on the Internet as an important reference,quietly linking the public reviews with the attractions,while most Internet users think that emojis and other sticker expressions can make up for the shortcomings of text expressions and better express emotions when expressing them,so emojis have become a nationally accepted online communication method in the mobile Internet era,and gradually become an indispensable communication element.In this thesis,a dataset of Chinese attraction evaluation text fused with emoticons is constructed,and the number of the dataset is 44,671.The main process of dataset processing is: firstly,the data are collected,then cleaned,deactivated words and meaningless emoticons are removed,some data of fused emoticons are artificially added,and finally,public release is made.The Chinese emotion classification dataset with fused expressions solves the problem that there is no public dataset for Chinese attraction evaluation texts.In this thesis,we propose the E2 G model to classify the dataset with high accuracy.The E2 G model preprocesses the text,then feeds into ERNIE-gram and Text GCN respectively.ERNI-gram is trained using its unique mask mechanism to get the final probability.Text GCN uses the dataset to construct a heterogeneous graph and uses documents and words as nodes to finally get a representation of nodes,and output probabilities,and compute the two probabilities to get the final result.In this thesis,to prove the effectiveness of the E2 G model,this thesis was compared with several groups of advanced models and classical models.After the experiments,it was shown that E2 G has a good classification effect on Chinese attraction evaluation,and the accuracy of classification is up to 97.37%,and compared with ERNIE-gram,Text GCN,E2 G accuracy is ahead by 1.37% and 1.35% respectively.Also to verify the performance of GCN and GAT on the dataset,two sets of comparison experiments were conducted,and the final results showed that ERNIE and ERNIE-gram combined with GCN and GAT,respectively,and GCN was 1.6% and 2.15% ahead in performance.Finally,to test the effectiveness of eight activation functions on the second layer of GCN,experimental comparisons were performed and the activation function with the best results,RELU6,was obtained.