The article that confused a consultancy firm with an author
The PDF of an article began with the following in the byline area:
Robert S. Bain, Steer Davis Gleave, Puerto Rico
What is interesting here is that the byline and the affiliation were contracted into one line in the PDF. This had gone wrong in processing, and since this is usually the place where authors are listed, "Steer Davis Gleave" (a consultancy firm, since renamed) and "Puerto Rico" had incorrectly been interpreted as co-authors.
How would this be handled in the canvas-plus-minigraphs model?
The canvas
The canvas would look exactly like what is shown above – no interpretation, just the text as it appears. The document layer's primary role is to define the narrative structure: telling us that what we see is a title, a byline, and an affiliation. That narrative structure is recorded as a minigraph asserting things about the text.
The canvas must represent exactly what is stated in the article. At most, some clean-up takes place – no interpretation. One natural temptation is to identify each author individually, but that is likely not the right level of granularity for the canvas. When I entered the world of SGML, there was an ongoing debate about what was "data" and what was "metadata". This unproductive debate is now replaced by the question of what is a "narrative" structure and what is interpretation.
The minigraphs
Three minigraphs linked to the byline line would indicate that, respectively, an occurrence of the entity Person, an occurrence of the entity Organisation, and an occurrence of the entity Country/Territory are found in that string. Those three minigraphs would be created by the entity-extraction capabilities of the Person and Organisation domains.
Thanks to their matching, linking, and clustering capabilities, those domains would also provide further minigraphs asserting that the detected Person occurrence matches a registered researcher entity, the detected Organisation occurrence matches a known organisation, and the detected territory links to a standard vocabulary. When our knowledge about the world changes, the relevant minigraphs get a new version.
We can also support other algorithms or subject-matter experts making assertions that all live side by side. We can pick and choose what is best in the consumption layer.
Why minigraphs must be mini
The term occurrence above is a key concept. We accumulate evidence about the existence of entities in the form of exhibits. The photo used in my earlier post illustrates the exhibit (the photograph itself) and the four occurrences of the Person entity – with the continued likelihood of creating a mismatch between the occurrence of my father and my own registered entity.
The idea is that there can and should be many of these minigraphs. minigraphs can be uniquely characterised by their type, provenance, and version. There can be multiple assertions of the same kind coming from different origins – or from the same origin with a newer version number. They can be stored as JSON-LD files, or collected in tabular Parquet files per type, or in many other ways. A main advantage of being mini is that they can be created independently. We avoid having to support multiple parties needing to edit one and the same XML file simultaneously.
Minigraphs can be visualised as little coloured tags stuck on top of the canvas near the place they are making an assertion about. They are more like tags than the XML tags of old: they point to a feature from outside it. They are also easier to version, replace, or retract than a section of inline XML.
Corrections are not exhibits
Sometimes corrections by a "human in the loop" are described as exhibits. But when we deal with corrections of entity extraction and entity matching, they are more like a mugshot: not evidence, but assertions in their own right. Rather than exhibits, we should see them as the output of human intelligence (HI) that runs in parallel to the AI of our algorithms. This needs to be captured accordingly – as a minigraph with its own provenance, not as a modification to an existing exhibit.