Research On Interactive Topic Modeling Method And Its Application

Online social network services(SNS)are changing the way people contact and producing a large amount of textual information.It provides important tools and opportunities for product marketing,knowledge management,and information dissemination.However,a large number of textual information has also brought huge challenges to the management of online social network platforms and marketing.Analyzing user behavior and modeling the content generated by users are of great significance in improving the management of the SNS platform,improving effectiveness of marketing,and promoting information dissemination.In this thesis,we propose two human-machine interaction strategies to incorporate human cognitions into topic modeling which are called interactive latent Dirichlet allocation(iLDA)and crowd based latent Dirichlet allocation(Crowd-LDA).The proposed iLDA model is a topic modeling method based on single-expert's knowledge.It assumes that the participant of the model is an expert user,which suggests that he/she is able to distinguish the improper words within a specific topic precisely and provide professional suggestions.We also present a new topic modeling algorithm that can incorporate the feedbacks of the expert effectively.Based on the scenarios where multiple users provide feedbacks,Crowd-LDA considers incorporating user knowledge from different backgrounds and a new topic modeling method based on crowdsourcing is studied.Finally,based on the two interactive topic modeling methods,we provide two different application researches which the first one is role discovery and the second one is service matching.The specific research content and contributions of this thesis include:(1)This thesis proposes iLDA to incorporate a single-expert's knowledge into topic modeling.In order to incorporate expert's knowledge,the proposed model provides several interactive strategies.Experts are allowed to tune the confusing words within a topic based on their own knowledge.Then the model will transform the results into a new topic-word distribution.Finally,a new topic modeling method that combines human's knowledge and data is proposed.The experimental results show that topic modeling incorporating human experts' knowledge outperforms than traditional methods.(2)Although topic modeling incorporating single expert's knowledge has better performance,it is not easy to find experts.To decrease the high cost of experts,this thesis presented another interactive topic modeling framework which based on crowdsourcing.It is called Crowd-LDA.When there are lots of common people who can give feedbacks easily,such as service matching and information retrieval,incorporating multi-users' knowledge is a better method.Crowd-LDA allows multi-person to give feedbacks for the results of the raw topics.The experimental results show that topic modeling incorporating multi common users' knowledge will also generate refined topics when experts are unavailable.(3)To explore the application of the proposed iLDA model,this thesis uses the model to extract user's topic preferences to discover roles in SNS.Then a new non-parametric method which combines user's topic preferences and behavior features is proposed.The model is able to automatically determine the number of roles based on data.(4)This thesis applies Crowd-LDA to match service in Internet of Things(IoT).Service matching is a common task in IoT service.Since each service request can be regarded as a feedback,service matching is a natural crowdsourcing system.Thus,we apply Crowd-LDA to extract topic signatures from service description text and match service based on these signatures.The experimental results show that Crowd-LDA can generate high qualified signatures.This thesis shows that topic modeling that incorporates human knowledge has great advantages.Interactive topic modeling methods can generate qualified topics which benefit for many tasks.