Text Sentiment Analysis Based On Deep Learning

Text sentiment analysis is an important task in natural language processing.With the development of Internet,a large number of text resources appear on the network,and the analysis of the emotions contained therein can extract great value and support the decision-making of institutions such as government and enterprises.We mainly focuses on two types of tasks in text sentiment analysis: sentiment analysis at the sentence level and target level.At the same time,we explore the methods and innovations of deep learning in text sentiment analysis task.The main work is as follows:1.In the sentence-level sentiment analysis,Long Short-Term Memory only extracts the serialized feature of texts,and lacks of consideration of the structural features.To solve this problem,Tree-Structured Long Short-Term Memory,a model that can extract structural features of texts,is combined with LSTM,and propose three kinds of new models: TC-LSTM,LT-LSTM and LCT-LSTM.2.In the target-level sentiment analysis,Long Short-Term Memory based models usually have complex structures,too many parameters,long training time,and can not represent the target words well.To solve those problems,we propose an new model named AB-TBSA based on attention structure,which uses the self-attention layer instead of Long Short-Term Memory for the text feature extraction,also,another attention layer is used to extract the target words' feature.The batch normalization layer is added for optimization,too.Finally,try to use the skip-gram model to train the position embedding of each word to add the position information to the attention model.Experiments on the sentence-level sentiment analysis task were performed on the SST dataset,the improved model achieved the accuracy of 88.7% and 51.0% respectively in the two-category and five-category tasks,both of them are better than the single Long Short-Term Memory or Tree-Structured Long Short-Term Memory model.Experiments show the possibility of combining the text's sequence features with structural features.Experiments on the target-level sentiment analysis task were performed on the Semeval-2014 dataset.The AB-TBSA-structure-based model achieve the accuracy of72.76% and 78.29% respectively on the Laptop and Restaurant subsets,which is better than Long Short-Term Memory based models in recent years.The analysis shows thesimplicity and efficiency of the AB-TBSA model.Visualization of attention weights also shows that AB-TBSA model can be well applied to the target-based text sentiment analysis tasks.Finally,experiments with the pre-trained positional embedding shows that they have a weak effect,but need further improving.