Aspect-level Sentiment Classification Based On Attention Mechanism And Graph Convolution

People can quickly and conveniently obtain information in their daily lives from the rapid development of the mobile Internet,the emergence of the current large network platforms is also changing people’s lives.Group buying platforms that facilitate people’s lives(such as Meituan,Word of Mouth),shopping platforms that bring convenience to people’s lives(such as Taobao,Pinduoduo),platforms that provide online communication and content sharing(such as We Chat,Weibo),and Social news media platforms(such as Surging news,Toutiao),commodity information,current affairs hotspots,real estate and finance-related topics all contain a lot of comments from people.These comment texts often carry the user’s own emotional color for such topics,and include the user’s subjective view.The comment information of these large Internet platforms has a certain role in improving the quality of good and guiding public opinion.Therefore,fully mining and using these comment texts has high social value.Aspect-based Sentiment Analysis(ABSA),a subtask under sentiment analysis,is a sentiment classification task that discriminates more specific attributes.When users comment on certain products or services,they may express different emotional tendencies toward different aspects of the product or service.Aspect-level sentiment analysis is to analyze the sentiment tendency of the user in a specific aspect.In recent years,deep neural networks have made great progress in the fields of image processing,network embedding,and speech recognition.Certain aspects of sentiment analysis can also learn from methods based on deep models.The main purpose of deep neural networks is to convert embedded word vectors into lower dimensions,explore the text features of long sequences,and then send the text vectors to the classifier for classification.In these thesis,we do the following work on sentence feature extraction and classification model of aspect level emotion classification:(1)Propose a new type of deep neural network model with hybrid extraction and multi-layer attention.Firstly,CNN and Bi-LSTM are used to extract the context short-distance and long-distance hidden features,and the two hidden layer vectors are directly spliced.This method is simple and practical.Then,based on the idea that words closer to the aspect word should be given a larger weight,the concept of position weight information is introduced,and position weight is added to the spliced feature vector.Then,the multi-layer Attention mechanism is used to automatically assign weights to the context feature vector,and the output vector of each layer of Attention is sent to the GRU network for non-linear integration,and finally sent to the Softmax classifier to obtain the final three-dimensional vector.The dimension with the largest value is the output predicted by our model(ie,positive,negative,and neutral).(2)Propose a classification model that explores the emotional dependence relationship between aspect words and aspect words by embedding coding and multi-layer graph convolutional network.First,use word embedding and Bi-LSTM to obtain the hidden features of context words and aspect words,and then perform syntactic parsing and coding based on aspect words on the context vector,combined with a two-way attention mechanism to simulate the relationship between each aspect word and its context word It means that.Finally,a multi-layer GCN is used after paying attention to the mechanical mechanism,the purpose is to capture the emotional dependence between different aspects in a sentence,and the results are spliced and sent to the output layer,and the emotional polarity is classified through the Softmax classifier.