Aspect-level Sentiment Analysis Of English Reviews Based On Deep Learning

With the development of information technology,many users like to post reviews of various aspects of products online.The goodness of the sentiment expressed in reviews has an important reference value for merchants and customers,so sentiment analysis of reviews is a worthy research direction.With the development of deep learning,many studies have achieved good results in the area of aspect-level sentiment analysis using deep learning,mainly because deep learning can effectively mine the semantic features of text and construct relationships between words.To address the problem that aspect-level sentiment analysis research lacks comprehensive consideration of lexicality,grammatical relation types,position,physical distance,and syntax,which leads to inadequate model learning,this thesis proposes a sentiment analysis model based on Feature enhancer Word vector and Mixed weight GCN(Feature enhancer Word vector and Mixed weight GCN,FWMGCN).Firstly,a new word vector is constructed by incorporating lexical feature vector,grammatical relation type feature vector and location feature vector on top of the word vector,which helps to enrich the feature information of the word vector.Then,physical distance weights and lexical weights are combined to reduce the noise and interference generated by the GCN model in the grammar learning process.Experiments were conducted on three public datasets with accuracy rates of 75.56%,81.87%,and 71.98%,respectively,indicating that the FWMGCN model can improve the accuracy of aspect-level sentiment analysis.In addition,to address the existing aspect-level sentiment analysis studies that only consider the grammatical structure of English comments and lack the research of combining grammatical distance and sentiment polarity knowledge on the grammatical structure,especially the lack of comprehensive consideration of sentiment polarity labels and sentiment polarity values,this thesis proposes a two-channel graph convolutional networkbased aspect-level sentiment analysis model(Two Channel GCN model,TCGCN).Firstly,the sentiment polarity structure graph and the grammatical distance weight enhancement graph are constructed,and the two graphs are fused by using the two-channel graph convolutional network,and finally,the multi-feature interactive attention method is used to mine the feature information of aspectual words.The experimental results show that the TCGCN model can preserve the complete feature information using the two-channel graph convolution,effectively achieves the fusion of the sentiment polarity structure graph and the grammar distance weight enhancement graph,and further improves the accuracy of sentiment classification compared with other single-channel graph convolution models.Applying the above research to a drug review aspect-level sentiment analysis system,the feature-enhanced word vector,mixed lexical-distance weight GCN and dual-channel GCN models can more accurately judge the sentiment polarity of drug reviews,which can guide users’ selection of drugs as well as help experts understand the strengths and weaknesses of drugs.