The term “vulgar linked data” comes from a 2022 article/blog post, Decentralized Infrastructure for (Neuro)science Or, Kill the Cloud in Your Mind by Jonny L. Saunders. It doesn’t seem to have caught on, but I think the idea behind it is worth engaging with critically. “Vulgar linked data” would be a more idiosyncratic, domain-specific form of linked data where it would be the responsibility of the users of data to map it to their own local ontologies rather than the creators’ responsibility to try to create it in a way such that these mappings could be created automatically. In Saunders’s words:

It is much more valuable to have low-barrier, vernacular expression usable by collections of subdisciplines and communities of people than a set of high-barrier, fixed, logically correct schemas. Researchers and people alike typically are only concerned with using the information within or a few hops outside of their local systems of meaning, so who is a totalizing database of everything for?

Saunders builds the idea of vulgar linked data from Lindsay Poirier’s description of “neat” and “scruffy” approaches to creating ontologies. On the “scruffy” approach, Poirier writes:

[…] for others on the working group, producing decidable results was not the most important aspect of the ontology language. Instead, this group was committed to a design logic that prioritized being able to express any set of statements - to “say anything about anything” - even if it may break some inferential properties of the language. The thought style that guided this community was much scruffier; it posited messiness and paradox as inevitable.

Saunders suggests that the academic community should embrace scruffiness, juxtaposing vulgar linked data with efforts to represent all data (or all data within a field) under a singular schema. I like this conceptually. I like the idea that data can be structured to meet the immediate needs of the people who need to use it and re-structured to meet the needs of others. It’s appealing because it puts the control over what is considered important to say in the hands of the people who are trying to say it. It seems more intuitive to explain your own way of thinking about something in your preferred manner than to try to adopt another person’s vocabulary in order to do it. It also goes against the grain of the way of thinking about data sharing I’m familiar with from Digital Humanities sources, both in terms of how it’s done and what it’s for. Saunder’s perspective is effectively the opposite of the one presented in the instructions given in The Programming Historian to prospective users of Linked Open Data. In that article, Jonathan Blaney writes:

LOD is structured information in a format meant for machines, and is thus not necessarily easy on the eye. Don’t be put off by this, as once you understand the principles, you can get a machine to do the reading for you.

If all datasets were openly published, and used the same format for structuring information, it would be possible to interrogate all of the datasets at once. Analysing huge volumes of data is potentially much more powerful than everyone using their own individual datasets dotted around the web in what are known as information silos. These interoperable datasets are what LOD practitioners are working towards.

In some ways, Saunders is writing from the future of what Blaney might hope would be constructed. Saunders works with medical data, where Linked Data approaches are more often used, and criticises some of the ways in which ‘interrogating all the datasets at once’ leads to poor outcomes. Saunders also goes into considerable detail on how deciding on any representation is an inherently ideological process, making a unified representation a complicated proposition. For Blaney, vocabulary reuse between datasets is a very important aspect of Linked Open Data, but he doesn’t consider the situation that Saunders describes where some institutions are able to specify schemas for representing semantic information that then become imposed worldviews for people seeking to publish information themselves. I would wager that in the field of history, things might play out differently because of the different demands and expectations placed on historians. The imposition of schemas would be less structurally enforced than in medical research, leading to a different sort of dynamic where the earliest participants would have the most influence on the semantic breadth of future ontologies. However, I think the impact would still be a chilling one.

All this said, I’m conflicted about Saunders’s technical proposals, because while they are systematic, they don’t address needs in isolation. Centralising formats means that we can’t express what we need to, so we’ll make format specification and translation between formats distributed. Distributing work means that work will (still) be duplicated, so we’ll build tools for sharing components that can fit into data transformation pipelines. The tools need to be hosted somewhere, so we’ll use peer-to-peer protocols. As a software engineer I’m not actually generally concerned about the technical feasibility of what Saunders suggests, because where it doesn’t work there are alternative ways forward. I’m concerned that the individual pieces don’t naturally lead to the kind of systemic change that Saunders is looking for and don’t achieve much in isolation. There’s a serious ‘chicken and egg’ situation: the existence of the technology requires a shift in priorities which requires the technology to exist to make following those priorities the path of least resistance.

There are other things that make me unsure. There’s an extent to which RDF can be used invisibly, ‘behind the scenes’, to manage things like identity and credit. But that is not what ‘vulgar linked data’ is intended to be for. It is there to present a structured account of what the people producing the data believe is “out there.” So it seems like to participate in this data sharing, it is not that all scientists have to become programmers (an expectation Saunders finds frustrating, as do I), but all scientists do have to become ontologists. I am uneasy about this because it seems to me like Saunders does believe that RDF can represent “anything about anything” at least in the context of science and that simpler, more humane user interfaces will allow us to represent what we need to. This paragraph is particularly telling:

RDF is highly polarizing, and many people have written it off as a lost cause because it is too complex. Much of the computing world runs off of table and relational databases rather than graphs. Though we tried to be careful to avoid endorsing any particular technology in favor of thinking about triplet links as such, the question of the literal implementation of the standards is an inevitable one.

There are reasons why graphs aren’t used for everything. In particular, they are actively bad for storing timeseries data: storing anything sequential in a graph is complex and retrieving it is incredibly inefficient. Effectively, you have to either store everything as a linked list that must be traversed fully to retrieve a specific node or connect a sortable value to every entity in the collection and then sort it after retrieval. It also means that some queries become computationally expensive in a way that is not immediately intuitive due to the time complexity of the algorithms that need to be used to complete the queries. Relational databases may be restrictive, but the restrictions they place on their users means that in general it’s possible to make a relational database that can respond to a query in a reasonable amount of time. While “neat” ways of building ontologies are generally associated with more totalising endeavours, another rationale behind the convoluted rules and aversion to “messiness” is to ensure that it is possible to extract information in a reasonable amount of time. By that, I don’t mean microseconds, I mean a guarantee that a query will complete in a time at all proportional to the amount of data needed to be processed, or even complete at all.

Of course, this is a surmountable problem: don’t use RDF (or any competing graph data format) where it’s going to cause grief. This is easier said than done, but Saunders’s idea of producing interfaces that allowed those resources to be referenced in RDF without subsuming their contents into it would probably need to be a long term plan rather than an initial concession. Though it seems somewhat antithetical to the premise, it also seems prudent to come up with restrictions and validators for “vulgar linked data” that make it more difficult to produce, for example, classes with inheritance cycles within them, and to ensure that basic search features work as intended.

For all Saunders talks about regarding social and technical problems as interrelated, the main thing I find frustrating about their proposal is a social one. It requires everyone to get together and decide they’re going to work towards a system that requires broad adoption to do anything at all. Assuming there was a collective push towards a system like the one Saunders describes, on the technical side there would be an initial period where there was significantly more work to do for everyone involved, involving a not-insignificant amount of trial and error. I’m not sure that the groups that would stand to benefit the most from this system and have the expertise needed to construct it are set up to be able to handle error. There may not be enough resources provided to them (in terms of time or money) to allow them to recover from those issues and build in a different direction. There may be pressure to present results as successes.

It’s because of these labour-related questions that I think it’s important to dig into what would be achieved and how minimal proofs of concept could be built demonstrating that the infrastructure would be useful. Some routes of enquiry could be: can research groups already using Linked Data approaches within the same subject area meaningfully bridge the conceptual gaps in their chosen frameworks? Is there a way to find out what a translation tool would need to do to be of assistance before going ahead and making one? I think that we are past the point where idealistic, overarching visions of what technology could do for us can be chased without knowing what the end results might be. There is so much inertia in the existing systems that it just isn’t enough to write a manifesto, even if that manifesto comes with an implementation guide.

I would really like for there to be constructive ways to approach works like Decentralized Infrastructure for (Neuro)science. What I would like is, if when reading proposals like this, there could be a discussion around it that takes apart the problem and solution and asks whether they naturally lead into one another. If there is something about it that could be adjusted to make it achievable. Whether there is any of it that could be done in a way that works towards the intended liberatory ends even if the new systems that people manage to build in the end are not quite the ones suggested. For all I’ve said in criticism, I think that Saunders is right to want the things that they want.