Text Sentiment Classification With Multiple Graph Convolutional Networks And MLP-Mixer

Sentiment analysis or opinion mining is the analysis of people’s perceptions,emotions,evaluations,and attitudes toward goods,services,organizations,individuals,social issues,events,topics,and their attributes.Since the beginning of the 2000 s,the text sentiment analysis task in natural language processing(Natural Language Processing,NLP)has become a very important research field,and it has also been further explored in web mining,text data mining,and information retrieval.Sentiment analysis has a wide range of applications,extending from management science to social fields such as marketing,finance,politics,media,and health sciences.The emergence of Internet social media(such as forums,blogs,Weibo,Twitter,etc.)has further developed sentiment analysis.Internet social platforms generate massive amounts of opinion data recorded in digital form all the time,and finding and monitoring websites and extracting opinion information from websites is still a tedious task.Because each site often contains a large amount of opinion text,the information in blogs and forum posts is difficult to decipher,making it difficult for ordinary readers to identify relevant sites and to extract and summarize opinions.Therefore,text sentiment analysis came into being.However,existing deep learning methods often cannot take into account the grammatical information,semantic information and context information of the text at the same time when processing text information.To further improve the performance of sentiment classifiers,the main contributions of this thesis are:(1)We propose an new attention mechanism WDA(Word and Dimension Attention)based on both words and dimensions.By selecting core words and extracting important dimension features of words,WDA can obtain richer textual feature information.(2)On the 190,000 Sogou news data set,WDA is combined with several different text classification models.The experimental results show that WDA has a significant effect on different text classification models.(3)Multiple Graph Convolutional networks(MGCN)is proposed,which simultaneously considers the complementarity between contextual information,syntactic structure and semantic relevance,and effectively integrates contextual information,syntactic information and semantics information.(4)The MLP-Mixer is innovatively introduced to mine the deep-level feature information and enhance long-range semantic dependencies in text,thereby improving the accuracy of text sentiment classification.