Research On Sentiment Analysis Method Based On Deep Neural Network And Transfer Learning

Sentiment analysis is widely used in social practice,such as lyric analysis,commodity evaluation analysis of e-commerce shopping websites,and evaluation and analysis of tourist attractions in tourism websites.Through sentiment analysis,the analyzed objects,such as the advantages and disadvantages of a certain commodity,attractions,etc.,can be presented to potential customers or tourists to help them make more comprehensive decisions.However,when using machine learning methods for sentiment analysis research,there are strict requirements for training data.First,training data and forecast data need to meet independent and identical distribution assumptions.Second,there must be a large number of labeled training data,and the categories should be balanced.However,data in the real world often has assumptions that do not satisfy independent and identical distribution,and the data categories are very uneven.In view of the above problems,this paper proposes two transfer learning methods based on deep neural networks to improve the performance of sentiment analysis.This paper proposes an sentiment classification framework based on sequence transfer learning method.It includes a downsampling method based on transfer learning and a cascade classification structure based on CNN.It can well solve the sentiment classification problem of imbalanced data sets.The traditional machine learning method uses the vector space model(VSM)to model the natural language.But the VSM makes the statement lose the order relationship and the context dependency,and also can not distinguish the words that are polysemy.Therefore,there are great difficulties in re-sampling and model learning.The experiment verifies that the algorithm framework proposed in this paper has 63%,64%and 63% accuracy,recall and f1 values in a few categories of extremely imbalanced realworld tourist attractions sentiment comment data sets.It is 12 and 5 percentage points higher than the models BalanceCascade and Multi-model Fusion,respectively.This paper proposes a multi-task learning model combining joint three-category and regression for the form of existing data labels.It learns sentiment strength values while learning emotional polarity,and makes full use of training data to improve the performance of sentiment analysis models.The model uses both character embedding and word embedding as input layers,using bidirectional LSTM and bidirectional GRU to extract semantic features.The self-attention mechanism is then used to redistribute the weight coefficients of the features,and the global maximum pooling operation is used to extract the maximum semantic signals for all features,and finally a fully connected layer is connected.Experiment results demonstrate that our approach's macro-average f1 value of the binary classification in the data set MinChnCorp is 94.48%,which is 3.5 and 2 percentage points higher than the multi-model fusion LR all and the deep convolution model CCB,respectively.