摘要：The overwhelming amount of web videos posted on the social mediawebsites make effective browsing and search a challenging task. Theuser-provided metadata, has been proved useful in large-scale videoorganization and retrieval. Search result clustering, which utilizesthe associated metadata to cluster the returned results intosemantic groups according to its involved subtopics, has shown itsadvantages. Most of the existing works on search result clusteringare devoted to solving the ambiguous problem resulted fromgeneral queries. In this paper, we propose the problem of\textit{faceted subtopic retrieval}, which focus on more complexqueries concerning political and social events or issues.Hierarchical topic model (hLDA) is adapted to exploit the intrinsictopic hierarchy inside the retrieved collections. Furthermore, thispaper offers a new perspective of multi-modal video analysis byexploring the pairwise visual cues deriving from duplicate detectionfor constrained topic modeling. We modify the standard hierarchicaltopic model by integrating: 1) query related Supervision knowledge(ShLDA) and 2) duplicate Relation constraints (RShLDA). Carefullydesigned experiments on web-scale video dataset validate theproposed method.
关键词：subtopic retrieval;multi-modal analysis;search result clustering;hierarchical topic model;social media;connected multimedia
