Research On Guiding Supervised Topic Modeling For Content Based Tag Recommendation

With the widespread of social media such like community QA sites and social blog system,many people can create richer and richer content information.But they also have to face the serious problem of information overload.Appropriate content tags can help users for better information retrieval.However,through statistics over 50%online content lacks tag information or even does not have tags at all.For most users,it's often painstaking and challenging to tag online content manually.On one hand,users often don't know what to tag and what tag to use for the lack of specific knowledge.On the other hand,the tags of complex online content labeled by users are often incomplete and inaccurate enough.Therefore,automatically recommending suitable tags for online content becomes a necessary task.At present,content based tag recommendation methods can be divided into two categories:one is the method based on keywords extraction,which assumes that tags are keywords.Such method requires much computation and is difficult to apply for the online tag recommendation;the other is the method based on topic modeling,which assumes that tags are latent topics.Such method ignores that tags are significant in their corresponding content,so that predic-tion accuracy is usually lower than the former method,but the efficiency is higher.On the basis of the latter research work,we further consider that the tags and their relevant words may have appeared(multiple times)in the corresponding content.We propose a novel topic model and present a general content based tag recommendation framework.Finally we implement a content based tag recommendation prototype system based on the model and the framework.Our main contributions are summarized as follows:1.We propose an automatic tag recommendation framework for online text editing system.This framework first acquires and preprocesses the raw data,and then uses these data to obtain the relevant words of tags based on word vectors,next uses the tags and their relevant words to train our topic model,and finally predicts the tags of new documents based on the model.The framework gives a general solution for providing personalized recommendation support for a variety of online content systems.2.We propose a generative model Sim Word extended from LLDA,which is guided by the tag and relevant word occurrence for content based tag recommendation.We also propose several special cases of Sim Word.Experimental evaluations on several data sets such as StackOverflow demonstrate the better effectiveness and efficiency of the proposed methods than similar work.3.We implement a tag recommendation system for StackOverflow based on the above technique and framework,which preliminarily verifies the effectiveness and ratio-nality of the technique and framework.