Aspect-Level Fine-Grained Sentiment Classification For Review Text Based On Deep Learning

With the rise of social networks,more and more people began to express their opinions and express their feelings on the portal.These text comments with personal subjective attitudes have great mining value.Text comments with a personal subjective attitude have great mining value.The traditional text sentiment classification mainly judges the overall emotion of the text,but due to the complexity of the text expression,a piece of text may contain multiple entities,and the emotional tendency of each entity is different.Therefore,the analysis of sentiment orientation of the text should be more detailed and more specific to each emotional entity in the sentence.With the rapid development of deep learning technology,it has been widely used in NLP tasks.This paper uses the deep learning technology as the basis for the level-level fine-grained sentiment classification research of text comments.First,for the problem that the up-sampling and down-sampling cannot solve the imbalance of the original data category,this paper proposes an algorithm to solve the aspectlevel fine-grained category imbalance--Batch Balance(BB).Through the analysis of the forward calculation of each layer of the neural network,the weights of the neural network are initialized: normal distribution and Kaiming distribution are combined.This solves the problem that the output disappears or explodes due to the neural network too deep.Secondly,on the basis of the original CNN and LSTM,this paper proposes an attention network that automatically focuses on different aspects of emotions--Attn-Bi-l CNN.Unlike traditional neural networks that transform text into a single semantic vector,the attention network transforms the text into a text matrix,where each row of the matrix represents different emotions in different aspects.Compared with the traditional neural network algorithm,it can achieve faster calculation speed.Finally,the BB algorithm and Attn-Bi-LCNN model in this paper are implemented specifically,and compared with the original neural network model,which proves the practicability of the model designed in this paper.