Research On Text Sentiment Analysis Based On Doc2vec And Deep Learning

Sentiment analysis,as one of the important components of natural language processing,is to identify and analyze the emotional polarity and intensity contained in the text.At present,sentiment analysis technology is widely used in tasks such as public opinion monitoring,financial analysis,and decision support,and has an important impact on government,finance,business and other fields.Previous studies have focused on the construction of emotional lexicon and the improvement of feature extraction methods,but the methods used are time-consuming and labor-intensive and poorly migrating.In recent years,deep learning has become the mainstream research method of natural language processing because of its advantages of automatic learning of data features.However,this method ignores the influence of word order information on classification results when dealing with sentiment analysis.Besides,there are deficiencies in extracting sentence structures and filtering invalid information.Therefore,solving the above problems by improving the deep learning model is of great significance for improving the performance of text sentiment analysis.Based on the above considerations,firstly,in order to solve the shortcomings of deep learning method of sentiment analysis in obtaining word order information and sentence structure information,this paper designs a convolutional neural network model CMCNN combined with Doc2 vec.On the one hand,as an additional vector of the convolutional neural network connection layer,the paragraph vector trained by Doc2 vec provides word order information that is lacking in the traditional word vector.On the other hand,the CNN model in this paper is different from the traditional CNN in the pooling layer.We use the Chunk Max Pooling strategy to ensure that the CNN model retains its location information while acquiring important features in the training process.So that the model can obtain both the semantic information that is effective for sentiment analysis and the sufficient sentence structure information.Then,based on the CMCNN model,this paper builds a fusion model At-CMC&BR that combines the advantages of CNN and RNN models.On the one hand,the model combines CMCNN's feature extraction ability from local integration to global,word order information extraction ability,sentence structure information extraction ability and BiRNN's context correlation information feature extraction ability.On the other hand,this paper adds the attention mechanism that can enhance the weight of key information in the model,which further improves the performance of the fusion model in solving sentence-level sentiment analysis.In the end,the proposed models are compared with several benchmark models.The results show that both CMCNN and At-CMC&BR have higher classification accuracy onthe IMDB and SST standard datasets,which verifies the validity of the two proposed models.