The Research Of Sentiment Analysis Based On LDA Model

In recent year,with the development of the Internet,the number of users of social software and e-commerce platforms is growing rapidly,which has generated a huge amount of textual data.It may carry lots of values and help people make decision correctly when mining and analyzing these texts.Sentiment Analysis is one of the most important methods of finding the sentiment tendency inside texts.And deep learning has been proven to have good performance on the task of sentiment analysis,where good vectorizations and high-quality classifiers are the main methods to optimize the neural model.However,if words have nothing to do with the sentiment,the two above optimization methods are inefficient,leading to unnecessary resource consumptions.To address this problem above,this paper focuses on feature engineering,studying on textual data to identify discriminative feature words and presenting two methods based on LDA(Latent Dirichlet allocation)model to improve the final classification accuracy.The main contributions of this paper are as follows:(1)A three-layer model based on LDA and TF-IDF is introduced,named TI-b LDAs model.This model combines advantages of both of them to identify words which have clear themes and high importance scores.(2)The c,v hyperparameters are presented,choosing the top-M words by their weights and then calculating a matrix sized N*K,using c,v as the pre-conditions to choose the feature words.The approaches we presented have been experimented on two different datasets IMDB and SST-2,and have been compared with four famous baseline models,and the results show that the feature words generated by our model are better than those by conventional feature engineering,leading to better accuracy.Therefore,the proposed method in this paper can have certain practical values.