Research On Sentiment Analysis Of Social Media Text Based On Graph Neural Network And Label Semantics

With the improvement of Internet infrastructure and the rapid increase of network users,social media has developed rapidly.Analyzing the online behavior of social media users helps to understand user needs,and has great commercial and social value.As a basic task in the field of natural language processing,text sentiment analysis aims to use computer technology to analyze and mine viewpoints,emotions,attitudes,and sentiment in text data.Graph neural network are widely used in computer vision,recommendation system and natural language processing because it can effectively process complex structure and obtain global information.In text sentiment analysis,graph neural networks can well model the paired relationships of words in complex texts by capturing the relationships between discontinuous words.Therefore,in recent years,a large number of research achievements in text sentiment analysis based on graph neural networks have been generated.However,text sentiment analysis based on graph neural networks still has some room for improvement.Firstly,in terms of text composition,optimizing text graph construction results can directly improve the performance of the model in sentiment analysis tasks.Secondly,data labels usually contain certain semantic information,and considering the semantics of these labels can reduce the prediction errors of the model.In particular,existing researches often ignore the inherent sequential nature of label in text sentiment ordered regression tasks.In response to the shortcomings of existing research,the main work of this article is as follows:(1)This paper proposes a new graph construction method: GC-KN.The core idea of this method is to find the key nodes in the graph,preserving the key nodes and their closely related neighbor nodes in the graph.Key nodes are extracted by an emotion dictionary or TF-IDF algorithm.Other neighbor nodes calculate their scores based on their word frequency and the number of key nodes in adjacent nodes,and retain the nodes with higher scores.(2)This paper proposes an LTC-DP model,which includes a data representation module,a label text association module,a deviation penalty module,and a classification prediction module.The data representation module is used to obtain a representation vector of the text.The label text association module fuses text information and label information by calculating the similarity between the text representation vector and the label representation vector,replacing the original One-hot representation of the label with a label distribution that blends the text representation and label representation.Due to the sequential nature of label in ordered regression tasks,the deviation penalty module penalizes the model according to the severity of different prediction errors.Finally,experimental results on six publicly available emotion classification datasets show that the GC-KN composition method proposed in this paper improves the classification performance of the graph neural network model by optimizing text graph construction results.This paper proposes an LTC-DP model to handle text-emotional ordered regression tasks.Experimental results on four emotionally ordered regression datasets show that fusing label semantics and considering label ordering can not only achieve better classification results,but also alleviate the severity of errors in incorrect classification results.