Context vs Content Metadata

In my continuing desire to analyze what types of metadata we use, and how we use it, I think there’s another useful distinction to be made, within the Added Metadata section. Long term readers will know I categorize metadata into six types:

1. Source – From the camera
2. Added – Added by people either real time or during editorial
3. Derived – the computer takes known information and derives additional useful information (GPS to address, for example)
4. Inferred – the computer infers additional information based on the available Source, Added and Derived.
5. Analytic – Using information in multiple images to infer metadata based on common image information
6. Transform – Any metadata that transforms the source – Color Look Up Tables, Motion Projects etc.

As we’re currently working on a new way of entering metadata in real time during a shoot, it became obvious to me that there are really two types of Added metadata: context and content. (Although both could be added automagically by some future computer program, for now they are human added.)

The real time logging app is focused on the who, where, when and what they’re doing. This is the context metadata. In our system it becomes both Keyword collections and is used for the Clip names. What it does not give me is the content of “interviews”. I include long takes of anyone talking in this category whether it’s been established as a formal interview or not. Here are the sound bites that will drive the story forward. This is Content Metadata.

I’ve found entering Content Metadata in FCP X to be easy, but still manual. In brief I use a common set of Keywords – Quote, Action, Problem – and enter details in the Notes field for each Keyword. This minimizes the number of Keyword Collections while providing the maximum amount of data entry, while also breaking that into findable bites. For Quotes I provide a transcript or summary of what is being said during that period. Action and Problem are valuable to isolate parts of longer takes, or to highlight issues that might make good drama.

My challenge now is to find a way or ways to automate the Content metadata, or simply its entry in the way that Lumberjack (our upcoming logging system) will do for Context metadata.  If we had good transcription technology available, we could use that and a keyword extraction technique. Right now, we don’t have a good transcription technology readily available. Although Nuance’s recent release of an API kit may be interesting, we are otherwise limited to Adobe’s Autonomy technology, which has shown less-than-stellar results. Apple may release a Dictation API for recorded files some time in the future (but that’s simply another way to the Nuance technology).

Probably our best bet for automating Content metadata is via Hollywood Tool’s VidKey.

By understanding the way we use “log notes” and the types of metadata that’s useful, we can then attack each step of the problem directly, finding the best tools for the job.