Research And Implementation On Sentiment And Topic Joint Model Based On Lifelong Learning

Topic Model(TM),as an unsupervised learning method,is widely used in the task of topic extraction.However,it is a data-driven model just based on word concurrence,which results in topics lacking interpretability and relevance.In order to solve this problem,many domainknowledge,such as sentimental lexicon,are applied to topic models.In the era of big data,the scale of data is getting larger and larger,the domain of data is more and more diverse,and the rate of data updating is more and more frequent.Especially in electronic commerce and microblog platform,great volume of review texts in multi-domains emerge as long as the trades are finished or hot events happen,which causes that the external static domain-knowledge is far from meeting the needs of mining high-quality topics.In the process of text topic extraction,in view of the fact that the external static domainknowledge cannot deal with the topic extraction task of updating frequent text,in this thesis,it is proposed that lifelong learning method is introduced to dynamically excavate the internal prior knowledge to improve the quality of topic mining.Meanwhile,in order to improve the learning efficiency of lifelong learning method,Variational Bayesian is combined with the knowledge regularization term to utilize the prior knowledge.The main contributions of this thesis are as follows:(1)Proposing a new topic model,Regularized Lifelong TM(RLTM),which adopts the lifelong learning method for mining internal prior knowledge,uses the VB method for solving parameters,and utilizes the self-learning regularization term for applying the internal prior knowledge.(2)Based on the RLTM model,proposing a new topic model,Regularized Lifelong Gaus-sian TM(RLGTM),which combines the external sentimental word vector knowledge with the internal prior knowledge,and utilizes the self-learning regularization term for applying the internal prior knowledge.(3)Based on the RLTM model,proposing a new joint sentiment topic model,Regularized JSTM(RLJSTM),which combines the external lexicon knowledge with the internal prior knowledge,and utilizes the self-learning regularization term for applying the internal prior knowledge.(4)Making a comparison between RLGTM and RLJSTM which represent different meth-ods of combine topic and sentiment,and doing contrastive experiments on the data from multi-domains to evaluate the topic quality and the model training efficiency.(5)Designing and implementing Text Sentiment Analysis System(TSAS)based on some models such as RLJSTM and so on.