More on tags, text, and categorization
In my previous post, I talked about our innate ability to aggregate things into useful groups. I also discussed how musical content didn’t even need to be characterized by a system in order for that system to make music recommendations. Finally I discussed how text is not suitable in and of itself as the basis for a correlation/organization between songs.
However, I’m not saying that text is always useless in the recommendation process of music. After all, there are many different motivations for listening to a piece of music, some of which are not based on a personal level of appreciation for the piece. For instance, say I’m interested in hearing Jazz from the 1940’s that has a bouncy, upbeat feel to it. Most of us have an idea of what that might sound like, and “bouncy” and “upbeat” seem to be good candidates for tags. However, you could write whole paragraphs on the music and still not really describe it well enough. It’s my opinion that trying to analyze the “content” of music solely through a tag/text based approach will not work.
The main problem is the act of representing a certain piece of content (music) in another medium (text). Performing an analysis of this translated content will never return a perfect characterization or correlation between two such translations. Frank Zappa’s quote mentioning “dancing about architecture” can serve as an example here: would you want to analyze dance moves to determine if two buildings are similar? Probably not, because there are an extremely broad assortment of dance styles, and many people (myself included) will really suck at reproducing any of them.
That’s not even the end of the problems for text and tags… because words themselves are not static, universally accepted, or discrete, there is a further added challenge to characterization in this fashion.
Tags are still a great way of “bootstrapping” the exploration process. Text is a pure symbolic representation, and even though it’s no stand in for the actual content being considered, it still can achieve passable recommendation results in some cases, such as when I want to hear music with the tags “acoustic” and “folk”. It’s results are also fairly good for tagging one’s own set of music (and where there is less conflict/overlap in how you are considering a given term). Plus, it’s kind of interesting to see what kind of content will show up under a “party” tag. Will it be the Back Street Boys or Bavarian Polka? Bon Jovi or Blondie? Black Sabbath or Black Eyed Peas? Given the right group of people, I think I could enjoy them all.