Research On Text Sentiment Analysis Based On Topic Model

All kinds of comments on internet is increasing and changing in an incredible speed.These reviews convey attitudes and thoughts of publishers.Thus these texts could be crucial in the areas of monitoring pubic opinions for the government,drawing up marketing strategies for enterprises,and making decisions when shopping online for individuals.Collecting,sorting,and analyzing these numerous reviews by hand not only cause a large consumption of money and energy,but also can not satisfy the demand of timeliness.That is why text sentiment analysis draws an increasing attention of both academics and industry researchers.The topic model is a generative probabilistic model which can extract latent semantic information of large-scale discrete data sets.In recent years,the topic model has been used widely in the fields like text classification,image classification,multi-document summarization,and recommendation system.The topic and sentiment unification model can efficiently detect topics and emotions for the given corpus and gets more and more attention of researchers.This thesis proposes two novel ways of topic and sentiment detection focused on long texts and short texts,which are Weighted Latent Dirichlet Allocation Algorithm(WLDA)and Biterm Joint Sentiment Topic Model(BJSTM).For long texts,faced with the low discriminability of topics in sentiment/topic analysis methods,this thesis proposes a novel way called WLDA which can acquire sentiments and topics without supervision.The model assigns different weights to different terms during Gibbs sampling,which enhances the impact of words which can convey the emotional attitude and reduce the effect of stop words in a broader sense.So the algorithm can get more discriminative topics.The experiments show that compared with Joint Sentiment/Topic model(JST),the classic model for conjoint analysis of topic and sentiment,WLDA not only has a higher iteration speed in sampling,but also gets better results in topic extraction and sentiment classification.2.For short texts,faced with the challenges in text analysis caused by its sparsity,there are just a few sentiment/topic analysis methods for short texts.This thesis proposes a novel way called BJSTM focused on short texts.A sentiment layer is added to Biterm Topic Model,thus a three-layer Bayesian model of “sentiment-topic-term” is formed.By sampling the sentiment and topic of each biterm,BJSTM builds the model on the corpus level.The model can depict the word co-occurrence of the whole corpus and use the abundant information of word frequency,which give the algorithm the ability of overcoming the sparsity of short texts to some extent.The experiments based on real comments on the Internet show that BJSTM gets better results in sentiment classification as well as topic extraction than JST and Short-text sentiment-topic model(SSTM).