text-driven vs. content-driven search

There have been a lot of observations and comparisons on the four media search engine: IBM, Columbia, YouTube, Google Video. My thoughts was not very different from those. Here comes my abstract comment on the two types of searching, namely, text-driven and content-driven search.

Why do text-driven search engines work?
People usually have a search goal in mind when go to a search engine. The goal is usually high-level, semantically constructed concept, i.e. what things are "about". An ideal search engine give things related to some specific concept, but also alternative things associated to the concept. For example, I want to search for clips about "tennis", I already have the concept about tennis, which can be specified by tennis games, tennis players, or associated to other sports--these are good alternatives in terms of semantic level. I don't expect clips like yellow or round objects that are visually similar because these are away from my search goal. Text like tags or metadata are information annotated by people to describe the media object (what the media are about), providing direct or indirect semantic links to certain concepts, which corresponds with most of the search goals.

What could content-based analysis help?
By content-based analysis I simply mean using some low-level features extracted from the media objects. Most studies use features derived from low-level signal processing and statistical analysis to infer high-level semantic meanings and to match the given query. Without certain domain knowledge, this attempt would be hardly competitive with text-driven search, since the semantics or higher-level structure of non-textual objects have tremendous variation. so sad? I'm thinking getting rid of the semantic level, maybe a much higher level information could be derived from such low-level features and help people expand or specify their queries. For example, people could get a sense about the styles of media from lower-level perception. One way to identify music style is based on its temporal structure. Therefore, it might be useful to detect the style or genre of the media objects, and use such information to represent or refine the queries. Maybe the music site pandora has built with similar idea?