Research On Sentiment Classification Method Based On Deep Learning

User review data directly expresses user attitudes and behavior preferences,and mining user review text can create new social value.At present,coarse-grained user comment mining methods have achieved excellent results.However,actual user comments contain complicated opinion information,and a large number of evaluation objects and attributes of evaluation objects are mixed.This study designed a better performance model for the problem of fine-grained sentiment analysis,which helps managers and researchers obtain more accurate user opinion information.This study treats the sentiment analysis problem as a common classification task to resolve.First,the defects of the recurrent neural network and its variants commonly used in sentiment classification tasks are analyzed.On the one hand,the ordinary recurrent network and its variants are distracting models;on the other hand,as the time step increases,the hidden state vector will be updated many times,which makes it difficult to save the information over a long distance.In order to overcome the defects,this paper proposes a multi-head attention memory network model.Secondly,it analyzes the limitations and weaknesses of the multi-head attention memory network in practical applications.On one side,the multi-head attention memory network can only handle explicit comment objects,but not implicitly expressed comment objects;on the other hand,the multi-head attention memory network belongs to a single-task learning model,which splits the possible correlation between comment objects during training Sex.Therefore,this paper redefined the sentiment analysis granularity and transformed the multi-class task model into a multi-task learning model.Finally,the multi-task learning classification model is used for example analysis,collecting user reviews of Dazhong.com,and discovering user expression trends and main concerns through word cloud analysis.Use coarse-grained sentiment analysis to find the time domain in which users' negative emotions are concentrated.Use the fine-grained sentiment analysis method to find the pain points and dissatisfaction in the user's consumption process,and put forward suggestions for improvement accordingly.The research results show that: the memory components in the multi-head attention memory network can be read repeatedly,which helps the model to capture information over long distances;the multi-head attention mechanism can calculate attention weights in multiple low-dimensional feature spaces to capture more abundant Semantic information;the location information of words and the distribution of data samples will affect the performance of the model.In addition,this paper compares six multi-task learning models with different structures,and takes the correct rate and Macro-F1 value as the measurement indicators,and believes that Transformer-Capsule network performance is the best.Transformer as the feature extractor in this paper has better performance than the commonly used Bi-LSTM and Bi-GRU.In terms of classifier,the Capsule model is better than the commonly used CNN + Softmax combination.The example analysis shows that the coarse-grained sentiment analysis method can provide early warning for managers,and the fine-grained sentiment analysis method can help managers find consumption pain points.