Research On Opinion Target Extraction Based On LDA Topic Model

With the rapid development of e-commerce,the subjective comments on the relevant products from customers grow rapidly.These reviews express peoples' all kinds of emotional colors and opinion polarity,including much valuable information.On the one hand,before buying the product,the latent customers often obtain the current product's information through the reviews of users who had purchased and decide whether to buy according to the information obtained.On the other hand,these reviews,for businessmen,play an important role in obtaining customers' feedback so that they can know which part of the product is attractive for customers and which part should be improved.So the merchants can maximize its commercial interests.Due to the mass information,it's almost impossible if using artificial way to handle these subjective reviews.So we need some techniques instead of artificial way.Sentiment analysis is born.And it has many important sub-tasks,one of them is opinion target extraction.The research of opinion target extraction mainly includes rule-based/template-based method and statistic-based method.Rule-based/template-based method requires domain experts to define the corresponding opinion targets and rules in the field,but it cannot meet the emerging new words and it is also poor portability and not cross-domain,and the opinion targets cannot be clustered.The LDA model is an unsupervised statistical model which is not only overcome the shortcomings of the above methods,but also need not require a lot of artificial markings,which has attracted the attention of researchers.However,as a bag-of-words model,the LDA model ignores the position information of the words and language structure information,so it is not suitable for extracting opinion targets.We need to extend the standard LDA to achieve the purpose of extraction.Among the extended models of LDA,there are many models to identify the opinion targets,but many of them cannot separate the opinion targets and opinion words.Zhao et al.proposed a MaxEnt-LDA model which introduces the maximum entropy model that can join the language features to make up for the shortcomings of standard LDA.It can identify the opinion targets and separate the opinion targets and opinion words.But the MaxEnt-LDA model still has some drawbacks.The MaxEnt-LDA model is characterized by lexical features,ignoring the syntactic features,and maximum entropy model has some drawbacks.To solve the above problems,a CLDA model which is the mixture of CRF and LDA is proposed in this paper to extract opinion targets.Firstly,the CRF is introduced into the LDA model to distinguish opinion targets,opinion words and background words.Secondly,opinion targets,opinion words and background words are divided into global and local conditions by adding the indicators.Global and local's distinction is that customers usually use frequent words in evaluating products which is easy to drown other opinion targets or opinion words,so the use of global and local is to distinguish between frequent and non-frequent words.The CLDA model can not only achieve the purpose of extracting opinion targets,but also separate opinion targets and opinion words.In order to verify the validity of the CLDA model,it is analyzed qualitatively and quantitatively using the review data set of the Restaurant field.In the quantitative analysis,CLDA is compared with the MaxEnt-LDA model and the experimental results show that the CLDA model has better performance in opinion target extraction.In the experiment,the feature selection problem in CRF is also discussed,and the validity of the selected word features,part of speech features and dependent syntax is verified by experiments.Finally,the research on Chinese opinion target extraction based on topic model is extremely rare.In my research,the CLDA model is used in extracting Chinese opinion targets and it is proved that CLDA model is also effective in Chinese datasets.