The content model I have long advocated organises scholarly content around two distinct layers: the canvas and the minigraphs.
The canvas
The canvas is an expression of the digital document, with as little interpretation as possible. This is a huge conceptual difference compared with traditional journal-article XML, in which we ignore the provenance of assertions and don't allow for multiple interpretations. The canvas contains markup and its suggested format is HTML5 – or more precisely, HTML5 plus embedded illustrations and media in a ZIP package.
Today, many secondary databases store records that have neither markup nor illustrations. That is wrong. Products not interested in markup can strip it out, or request it to be stripped out. But the content platform should not make that decision on behalf of all possible consumers.
The minigraphs
The minigraphs are annotations layered on top of the canvas. With "annotations" we mean any assertion we want to make about the canvas, or portions of the canvas. The word "graph" indicates a collective of such assertions expressed in RDF – for instance JSON-LD. The prefix "mini" emphasises that the annotations are as single-minded as possible: they record the outcome of one process at a time and therefore have a unique provenance. There can be minigraphs about the same assertion but from different origins, and the minigraphs can contradict each other. Through the knowledge layer, the standard view on each entity is then assembled from all available minigraphs.
With "as little interpretation as possible" we mean that the canvas must remain neutral. It detects person names but refrains from calling these "authors". The fact that the names represent a contributor occurrence, and more precisely an author occurrence, is the result of a distinct interpretation step recorded as a separate minigraph.
One way to visualise the canvas with minigraphs is as a Word file with hand-drawn annotations stuck on top – the way a copy editor would add them on paper in the old days. But the annotations are not added directly on the page; they exist in their own layers, like layers in Photoshop, and we can pick and choose the layers that are relevant. This also means we can have multiple, independent annotators working on the same canvas at the same time, each contributing their own minigraph.
The annotations are more truly "tags" than the XML tags of old: they are stuck onto a feature and say something about it in an appropriate vocabulary. Depending on the target audience the vocabulary is more or less detailed. If one wants to tell the outside world about it – where "outside world" means machines interpreting the data in the spirit of the Semantic Web – one might choose a widely understood external vocabulary. Deeper annotation, needed inside a data platform, can use a domain's own vocabulary. We do not need to burden the outside world with fine-grained internal detail unless it is really needed.
The annotations also allow us to support other algorithms or subject-matter experts making assertions that live side by side. We can pick and choose what is best in the consumption layer, taking note of entitlements.
This is a huge difference compared with the old approach of trying to capture everything in one XML file. We must avoid the old-style desire to capture each detail in one big file – it doesn't matter whether that file is XML or JSON – and instead think carefully about who we tell what, and why, and separate interpretations accordingly.
Three ground rules for the data
In the annotations, we want to adhere to three principles:
- Record provenance. Always make it clear who did what and why.
- Ground data in time. Many properties are grounded in time – countries cease to exist, organisations evolve, people change names – but this is seldom modelled properly.
- Respect zero. There is a meaningful difference between not recording anything and recording that something isn't there.
These three rules sound simple. Violating them – and legacy data systems violate them constantly – creates compounding problems that are surprisingly expensive to fix.