Retrieving what's relevant in audio and video: statistics and linguistics in combination