Text Sentiment Classification Based On Attention Mechanism And Fusion Of Neural Networks

With the developments of the emerging technologies like artificial intelligence and big data,the trend of the intelligence and big data developed more rapidly.And it will create a lot of value if we apply this kind of technology to all aspects of society.People have been familiar to express opinions through the Internet.And we can obtain effective emotional and semantic information in massive network evaluation texts through the deep learning method of sentiment analysis.So that many social networks platforms,e-commerce company,and government agencies can effectively provide humanized services such as news recommendation,shopping navigation,and public opinion surveys to the society.Therefore,it is necessary to construct the network model by proposing effective sentiment analysis methods to ensure the accuracy and utility of text sentiment analysis.(1)Aiming at the issue that traditional Chinese evaluation sentiment classification pays less attention to deep emotional semantic information,this thesis proposes a method named multiple attention feature fusion and the sentiment classification model MTA-CBG(Multi-Attention Convolution-Bi GRU)is constructed.Traditional word vectors cannot solve the polysemy of Chinese text effectively,so this thesis correlates the features between words by the construction of a self-attention word vector matrix.Also,a Multi-scale Wide Convolution(MWC)structure is constructed for comprehensive extraction of the local features and integrates the two types of features.The text serialization features are learned through Bidirectional Gated Recurrent Unit(Bi GRU),which can solve the long-distance dependence problem while extracting a wider range of text features.Finally,the sentence-level features are associated by the constructed Attention-Highway layer to extract the deep semantic and emotional features of the text.The comparative analysis of multiple sets of experiments confirm that the proposed method and the constructed model in this thesis have improved the F1 value and accuracy of Chinese evaluation text sentiment classification.(2)Aiming at the issue that the traditional method of the text vector representation cannot fully represent the emotional semantic information contained in the Chinese text when processing the sentiment classification task.,a Chinese sentiment classification model named Multi-Granularity Convolution Capsule Network Model(MGC-CAP)is constructed.In sentiment classification task,the text vector representation training by the language model is usually used as the input of the model.Aiming at the problem that single language model cannot fully abstract the characteristics of Chinese evaluation text,a textual multi-granularity representation method based on pre-trained language model BERT(Bidirectional Encoder Representations from Transformers)is proposed.In this thesis,the coarse-grained word vector of Word2 vec is trained based on the traditional distributed representation method,and the fine-grained character vector is output by fine-tuning the pre-trained BERT model.The different granularity vector representations of the text are input into the improved multiple channel convolution capsule neural network after the embedding layer.The traditional convolutional neural network uses the maximum pooling strategy to further extract features.However,this operation may ignore other important features of the text and only select the most significant features caused by the loss of some deep emotional semantic information in the text.Therefore,for the purpose of construting the overall emotional semantics,this thesis constructs a capsule layer after the convolution layer to associate the text features in low layer and the features in high layer,and uses the dynamic routing algorithm to reasonably screen the feature information Experiments were conducted on multiple sets of Chinese evaluation data sets to affirm the utility of the improved method and model constructed in this thesis on sentiment classification task.