Study of cross-media topic analysis based on visual topic model

Research on cross-media topic analysis methods, which utilize semantic of multimedia data to describe topics of cross-media documents. As the emerge of food safety related multimedia data, topic analysis based on single media data can't obtain full topics, causing the problem of inadequacy of semantic. A cross-media topic analysis framework is proposed in this paper. Firstly, generative methods are used to get semantic of text and image data respectively. Then a visual topic learning algorithm is presented to construct visual topic model and map visual data to text topics. This method can solve the problem of consistent semantic description of cross-media data. On this basis, food safety topic tracking is achieved and experiment results also show its effectiveness.

1 INTRODUCTION

With the rapid development of Social Network Service (SNS) applications, Internet has become an important channel for people to get information and to express their opinions. Because of a wide variety of multimedia data in SNS ^[1], the fusion of all kinds of media form information from different sources using a cross-media data mining methods is very important for topic analysis and hot topic tracking, which is also an important issue to monitor emergency events in complex decision-making environment . It's needed to adopt a cross-media mapping method to extract data from all possible media, converted to a unified semantic space for processing. In this way, more knowledge reflecting current hot events can be discovered, and the status of the events can be accurately assessed too.