onlin multiscal dynam topic model tomoharu iwata takeshi yamada yasushi sakurai naonori ueda propos onlin topic model sequenti analyz time evolut topic document collect topic natur evolv multipl timescal word consist hundr year word emerg disappear period dai propos model current topic specif distribut word assum gener base multiscal word distribut previou epoch long timescal depend short timescal depend yield robust model deriv effici onlin infer procedur base stochast em algorithm model sequenti updat newli obtain data mean past data requir make infer demonstr effect propos method term predict perform comput effici examin collect real document timestamp 