Research On Aspect-Level Sentiment Classification Based On Syntactic And Semantic Information

Aspect-level sentiment classification is an important task in the field of text sentiment analysis,which aims to predict the sentiment polarity of each aspect word in a given sentence.With the popularity of e-commerce and social platforms,an increasing number of text comments with emotional tendencies are emerging,bringing significant research value to both business and scientific fields.By conducting sentiment analysis on these texts,consumers can be provided with decision-making support,and feedback information can be obtained to timely upgrade and improve products for businesses.However,existing aspectlevel sentiment classification models often do not fully integrate syntactic and semantic information of sentences when predicting aspect sentiment polarity.This leads to the neglect of sentiment words related to the aspect and results in incorrect predictions of the sentiment polarity.Therefore,this thesis proposes a method that combines syntactic and semantic information to predict the sentiment polarity of aspect words.The main work of this paper includes:(1)An aspect-level sentiment classification model based on a multi-view gated graph convolutional network is proposed.In order to better handle data with less obvious syntactic features,the model uses gated graph convolutions to integrate syntactic and semantic information,thus allowing the model to more reasonably utilize syntactic and semantic information to make decisions.In addition,sentiment knowledge and aspect perception information are added to the dependency relations to redefine the syntactic dependency graph,enhancing the representation of sentiment features between aspect words and sentiment words.For sentences with multiple aspect words,the model captures the emotional dependency information between aspects as auxiliary information for specific aspect emotional features.Finally,the model adaptively integrates information from multiple views through a multi-layer residual perception network.Experimental results on four public datasets show that the model achieves good sentiment classification performance.(2)An aspect-level sentiment classification model based on selective attention graph convolutional network and aspect prompts is proposed.The model aims to address the issues of insufficient aspect sentiment information mining and noise information introduction in the use of too many graph convolution layers.To effectively mine the aspect semantic information of sentiment,the model utilizes the masked language modeling task in pretrained language models to obtain the sentiment prompt information implicitly hidden in aspect words.Meanwhile,in order to obtain important context information related to aspect words and reduce the impact of noise information on the model,the model uses a Selective Attention Graph Convolutional Network to capture dependency structure information and important information on sentence sequences,selects k context words with the highest attention scores to locate the aspect sentiment words,and obtains a matrix of selective attention scores.Then,another graph convolutional layer is applied to extract important semantic information on the matrix of selective attention scores.Finally,the aspect prompt information and the graph convolutional selection information are integrated through an attention mechanism.Experimental results on four public datasets demonstrate that the model has a positive effect on predicting aspect-level sentiment polarity.