A Research Of Text Sentiment Classification Algorithm Based On Attention Mechanism

In the research of natural language processing(NLP),text sentiment classification is a study hotspot.With the explosive growth of data in the information age,a large quantity of text data generated from users has been amassed when utilizing the tools of Internet.How to extract users'emotional tendency information effectively from these texts presents a challenge to researchers of NLP today.The text sentiment classification methods,as an effective choice to slove this problem,have been studied broadly by researchers.Existing text sentiment classification methods mainly use technology of neural net-work to finish feature extraction from the raw data,and then classify the texts based on the extracted features.Therefore,how to extract more efficient features from the original text has become one of the main research directions in the current text sentiment classi-fication research.Currently,the existing main problems in text sentiment classification models are as follows:First,in document-level sentiment classification tasks,they usu-ally focus on the overall characteristics of the text,and ignore the semantic relationship and position dependence in the text,which makes the model cannot deal with word order changes in the prediction process;secondly,many models in the document-level senti-ment classification tasks did not take into account the effectual combination of local and global features,which affects the actual model performance;thirdly,in the aspect-level text sentiment classification task,many studies have insufficiently explored the relevance of the context,which leads the problem of feature modeling;finally,existing methods in the aspect-level task are unable to integrate the target aspect feature with the context fea-ture very well,resulting the problem of incomplete feature structure.For solving the insufficiencies described above,the contents of this thesis are listed as follows:(1)Conduct extensive research on the existing research of text sentiment classifica-tion tasks about document-level and aspect-level tasks,study the existing main methods and technical means for solving such problems,and summarize the deficiencies in the research and the areas to be improved.(2)Propose Convolutional Neural Network with Attention-based Bidirectional Gated Recurrent Units(AttBiGRU-CNN).In the existing document-level sentiment classifica-tion research,there are problems of insufficient modeling of the semantic relationship and position dependence of text information,and the combination of global features and local features.Therefore,the model in this thesis combines the idea of AttBiGRU with convolu-tional neural network,and pays attention to the feature modeling of semantic relationship and position dependence as well,which increases the performance in this classification task.(3)Propose Attention-based Bidirectional Unlinear Graph Convolution Network(Att-BULGCN).In many existing researches on aspect-level text sentiment classification tasks,there are problems of modeling of contextual relational features and insufficient fusion of contextual information and aspect features.Therefore,the proposed model combines the uninear weight graph convolutional network for deep mining of the contextual feature relationship,and uses the attention mechanism for the fusion between features,which provides better performance in the aspect-level classification scene.(4)For the above two proposed models,this thesis carries out the experiments on several frequently-used datasets.The outcome of experiment indicate that the proposed models in this thesis outperform the other contrastive models in the corresponding experi-mental indicators.The experimental results show the effectiveness of the model proposed in this thesis.