Author's Sentiment Analysis Based On Probabilistic Topic Model

In recent years,probabilistic topic model is widely used in the field of text analysis.Topic model for text analysis shows several advantages.Firstly,topic model has a good foundation of mathematics and scalability.Secondly,compared with VSM model,topic model can gain lower text dimension and computation complexity in the process of modeling.The last,the method based on topic model is unsupervised in general,and it doesn't rely on artificial semantic dictionaries or sentiment dictionaries.The topic model for text sentiment analysis is called sentiment-topic model.The existing sentiment-topic model's generation process is a three 'layers generation process:document-topic/sentiment-word.The model can discover the topic information and sentiment information in documents.In fact,most of the text contains the author information,and analysis the author information is very valuable.The sentiment topic model do not model the author,so we can not analysis the author's sentiment.In this paper,the contribution made by the following:1.We propose a probabilistic modeling framework,called Author-Topic-Sentiment Mixture(ATSM)model,which based on Latent Dirichlet Allocation(LDA)to include authorship information and sentiments information.The proposed model can reveal the sentiment-topic and author's sentiment.We have three parameters to estimate:a distribution over the author's topics,a distribution over author's sentiment with given topic,and a distribution over sentiment-topic associated with words.We use Bayesian estimation and Gibbs Sampling to estimate parameters.The model can not only find the semantic information and sentiment information in text,but also find the author's topic and sentiment information.2.This paper presents a framework for the analysis of authors which are modeled with ATSM.The result is a two-tier clustering structure,including the outer topic class and inner sentiment class.Topic class is based on the author's topic was distributed clustering;sentiment class is the subject of further class clustering obtained by the sentiment similarity.This paper shows the physical meaning of the three parameters in ATSM model,and evaluates the generalization ability of the model with Perplexity.Experiments show that ATSM model has good generalization ability.This paper also analyzes the clustering results with ATSM model-based clustering framework,and shows the author's characteristics in deferent topic class and sentiment class.Experiments show that the authors in same topic class but in different sentiment class are similar in topic but different in sentiment and the authors in same sentiment class are similar in both sentiment and topic.