Research On Text Emotion Classification Algorithm Based On Deep Learning Technology

With the Internet technology and network platform is becoming more and more popular,more people are involved in the network information interaction.A large number of text information emerged in blowout now.People express their views through the network and social platform.By digging through these text content,you can not only understand user preferences,understand social hot issues,but also get information about product and service evaluations,etc.So text sentiment analysis has its importance both from an economic and political perspective.Through the analysis,processing,induction and reasoning of these texts,we can find great commercial value in the application of network public opinion discovery and other applications.Therefore,this paper studies the text sentiment classification algorithm based on deep learning technology,and the specific content includes the following aspects :First,a CNN-BiLSTM network model was implemented.This model combines the advantages of Convolutional Neural Networks(CNN)that can extract high-dimensional text features and Bi-directional Long Short-Term Memory(BiLSTM),which is good at dealing with sequential problems.This paper studies the model on text sentiment classification: including designing convolution kernels of multiple sizes to extract features at different latitudes,and using a KMax pooling layer that can retain multiple strong features instead of only those that retain the strongest features 1Max pooling layer.Experimental results show that the CNN-BiLSTM-KMax model performs better on text sentiment classification than other traditional network models.Secondly,the attention mechanism was introduced into the CNN-BiLSTM-KMax model,and a new ACBiLSTM-KMax model was proposed.Focusing on keywords with significant influence on the sentiment classification task can more accurately analyze the sentiment of text.The experimental results show that after the Attention mechanism is introduced into the model,the processing effect of the model is indeed improved.Then,based on ACBiLSTM-KMax,this paper uses a bidirectional language model to train the ELMo model of dynamic word vectors instead of the commonly used Word2 vec model.The experimental results show that the ELMo model can effectively improve the quality of word vectors based on the grammatical transformation of words and polysemy,so that the model can be optimized to a large extent.Finally,this paper verifies and explores the relationship between various parameters in the model proposed in the paper and experimental results,including Dropout value,Epoch value,word vector quality,and Batchsize parameters.The experimental results show that the quality of the model is positively related to the quality of the word vector.An excessively large Dropout will reduce the model's effect,while a too small value will not alleviate the over-fitting situation.The effect of the model will increase first with the number of trainings.After reducing it,it finally entered the overfitting state.When considering the Batchsize value,this paper must fully consider the time cost and training effect during training,and then combine the samples to make a choice.