Research On Text Sentiment Analysis Based On Deep Learning

Text sentiment analysis,as one of research fields in natural language processing,is of great significance in distinguishing text sentiment from different granularity.To this end,this thesis studies the shortcomings of the existing models in different sub tasks of text sentiment analysis.The specific study is as follows.1.Considering only semantic embedding is used in the existing text sentiment analysis model,and convolutional neural networks(CNN)can only extract local features of the text as well as the fact that word embedding and convolution neural network are the key technologies in sentence-level sentiment analysis task,semantic embedding,sentiment embedding and lexicon embedding are applied for text encoding,and word ambiguity is eliminated by extracting word context.On the basis of CNN model,three attention mechanisms including LSTM attention,attention pooling and attention vector are used to extract global features of the text to build attention CNN models.Additionally,cross-modality consistent regression(CCR)and transfer learning are introduced into this task for the first time to improve the model performance.The models are verified on several datasets,and the comparison results show the effectiveness of the models.2.Aiming at the problem that the existing models can only fulfil targeted sentiment classification task in the simple sentences,the new dynamic memory network,in which the input module,question module and memory module are optimally designed,is established to model the targeted sentiment classification task into question answering system for the first time.Position information and skip connection are considered in the input module to build rich representations over the sentences.To deal with target composed of multiple words,designed sentiment question about target is encoded by a GRU in the question module instead of averaging target word vectors.As for the memory module,on the one hand,the introduction of attention based GRU network and inner attention GRU network eliminates the weight bias caused by original soft attention in extracting memory information at each attention step.On the other hand,memory updating mechanism can distill sentiment-relevant features of the target from multiple attention memory information.Experimental results show that the proposed models can achieve the ideal results on several datasets,and demonstrate that the models can identify target sentiment in the complex sentences.