Research On Aspect-level Sentiment Analysis Based On Pre-training Model And Graph Convolutional Network

As a medium of communication between people,language plays a very important role in daily life.Text,another form of language,is one of the most common and abundant unstructured data on the web.Aspect-level sentiment analysis of text data,fine-grained exploration of people’s subjective emotions,and understanding of people’s opinions or insights are of great significance in business,society,public opinion surveys,and human-computer interaction.Aspect-level sentiment analysis aims to classify the sentiment of different aspects of a sentence.This kind of fine-grained sentiment analysis can reveal the sentiment tendency in text data in a more specific and detailed manner.In this paper,it is difficult to effectively distinguish different aspects of a sentence in aspect-level sentiment analysis tasks,and the accuracy of sentiment classification is not high.Make full use of the current popular preprocessing language model,graph convolution technology and dependency syntax analysis,etc.,and build an aspect-level sentiment analysis model through deep learning technology.The specific research content of this paper is as follows:(1)This paper establishes an aspect-level sentiment analysis model based on XLNet-LCF.The model uses the XLNet pre-training model to obtain contextual semantic features bidirectionally,introduces the context focus mechanism(LCF),and captures the local context of the context with aspect words as the focus,which can effectively reduce the problem of mutual interference between emotional words of different aspects;combined with multi-head self-attention The force mechanism deeply extracts the semantic feature structure in the global context,constructs the emotional weight matrix,and normalizes the matrix to improve the training speed;finally,input the emotional analysis layer to judge the emotional polarity.The accuracy and F1 values of the model in Laptop,Restaurant and Twitter datasets are 80.12%and 76.59%,85.24%and 76.90%,81.20%and 73.37%respectively,which effectively verifies that the model is progressiveness and effective in aspect level emotion classification tasks.(2)User comments are generally made from multiple angles,which leads to multiple aspects in a sentence,and longer comments usually contain more effective information.This will lead to more complex sentence dependencies parsed,and a large number of parameters will be generated in the graph convolution process,which will cause problems such as noise,or lead to model training overfitting,etc.Therefore,it is necessary to choose information and leave information that is effective for sentiment analysis.For this reason,this paper proposes an improved dependency graph convolutional neural network joint model,which improves the dependency tree and filters the sentences in sentences.Effective information,choose the aspect words as the center,cut out some dependencies,and build multiple screening matrices and vectors,so that the final weight matrix is centered on the aspect words,so as to realize the sentiment analysis of multiple aspect words in a sentence Only target dependencies are preserved.In order to verify the effectiveness of the model proposed in this article,a more challenging MAMS dataset was used to verify its accuracy and F1 values of 87.27%and 86.08%,respectively.The experimental results showed that the model outperformed the comparative model.