The Research Of Chinese Sentiment Classification Method Based On Topic Model

With the evolutionary development of the "Internet +”,all aspects of people's living is gradually filled with network,more and more people express their views in the Internet.Subjectivity documents about a movie,a commodity,a hot topic of the society and so on are inundated with the entire Internet.Researching these subjective documents' s emotional tendencies is of great practical significance about analyzing consumers' attitudes and emotional changes of the public.Therefore sentiment classification of natural language has been a hot issue in the field of document processing.This paper has done the following research out of the great theoretical and practical significance of sentiment classification.Firstly,this paper introduces the main technique used in researching sentiment classification and it also systematically expounds the evolution of document representation model.From the original VSM model to latent semantic-based LSA model and PLSA model,this paper focuses Latent Dirichlet Allocation model,including principle,generation process and parameter derivation,as well as some specific application in the emotional classification.From this it's feasible and effective to introduce LDA model to sentiment classification.Secondly,this paper presents a topic-weighted LAD model of sentiment classification to solve differences about topic feature quality in LDA model.This paper introduces a neural network based method to get distributed representation,with which to calculate the similarity of words and the internal similarity of topics,then transfer it to the weights on the topic dimensions.And take the weighted document-topic distribution to train classification.The following experiment shows that the 5% improvement of original LDA model's classification which proves effectiveness of this paper.Finally,this paper seeks an unsupervised sentiment classification method to reduce labour marked cost.With the help of HOWNET to calculate words' emotion value,It can get document-topic distribution and topic-word distribution with no supervision by the LDA model's contribution.Through words' emotion value with high probability,calculating the emotional distribution of topic.And further calculate the emotional distribution of document with the topic-wighted algorithm.The following experiment is entirely unsupervised and leads to good result,so the paper has some research significance.