Skip to content

A grammar for how knowledge connects

Twelve kinds of link in four families, with their properties and the rules that follow from them. Why a search must rank what a statement rests on above what it merely resembles, worked through on a public guideline on hand hygiene in district nursing.

Written by

Essay8 min read

What I wrote down today is called, in its formal version, an ontology. Grammar is the better word: it is not about the content, but about the rules by which you put content together.

It began with something that half failed. Last night I set out to build a knowledge graph: sources as points, with the links between them. It went less well than I expected, and in the end fairly well after all. This evening I sent the first clustering to the content expert I work with, and added that the clustering was probably completely wrong, because I lack the domain knowledge.

I meant it. I can build a graph. Only someone who knows the field can say which link lies between two documents. So I had to give the field a language in which to correct me.

At 18:21 I sent my own summary, four pages in mathematical notation. Five minutes later I withdrew it: it was noise. At 18:43 the same grammar followed in plain Dutch, also four pages, with one request: “I make quite a few assertions and set out postulates, implicitly and brazenly, as truths, but only to get the ball rolling; so please shoot down my wrong assertions without mercy.”

Below I explain what it says, and why care needs it.

Resembling a source is not resting on it

A search system that works on similarity turns every passage into numbers and looks for the passages whose numbers lie closest to those of the question. A language model is good at this. It finds what resembles the question, even when the words differ.

But resembling is not the same as resting on. A passage in the same words as the question can be about a different situation, or come from a version that no longer holds. A passage in other words can be the very rule the answer rests on. To an index built on similarity, every passage is a loose piece of text. It does not know that one document stands on another. Searching along a graph works differently: it knows that knowledge has structure.

That structure is usually in the documents already. A protocol names the guideline it is based on. A work agreement names the protocol it follows. What was missing was a fixed way of saying what kind of link that is.

The grammar has four families of three kinds of link each. For each family, one question decides whether a link belongs to it. You ask the questions in this order.

  • Origin: derives from, extends, specialises. The question: would one source exist if the other had not come first?
  • Basis: rests on, informs, confirms. The question: what happens to one source if the other proves wrong? If it falls, it rests on the other. If it weakens, the other informs it. If it stands but is less certain, the other confirms it.
  • Structure: contains, complements, integrates. The question: can the whole be understood from its parts, or the parts only from the whole?
  • Practice: applies, puts into practice, enables. The question: does the link make something general concrete, or something possible real?

The rules

The links have properties, and rules follow from them. Five matter to anyone who searches.

  1. Every link has a direction. A protocol rests on a guideline, not the other way round. Turn the direction around and you draw the wrong conclusions.
  2. Some links carry through. If A rests on B and B rests on C, then A rests on C too. That holds for “rests on”, “derives from” and “extends”. It does not hold for “informs”: if A informs B and B informs C, nothing follows about A and C.
  3. What carries a source also carries whatever derives from it. If a protocol rests on a guideline, and a team’s work agreement derives from that protocol, then the work agreement rests on the guideline too.
  4. If one source complements another, that works both ways, and the two belong together in the answer. Ask about one and you get the other with it.
  5. In a search, what counts is the distance along the links. “Rests on” is closest, then “derives from”, then “informs”, then “complements”. So what a statement rests on counts for more than what only informs it. What merely resembles the question is not a link at all. It is a likeness of words.

A public guideline

Take the guideline Basishygiëne Wijkverpleging (basic hygiene in district nursing) from RIVM, the Dutch National Institute for Public Health and the Environment, authorised in August 2024 and online since 27 September 2024. It spells out how it connects to other sources, and that makes it a good example.

It says what it rests on: “In content, this guideline is based on the SRI guidelines below.” There are six of them, among them the one on hand hygiene from 2023. It says what it replaces: the advice for home care from the LCHV (Landelijk Centrum Hygiëne en Veiligheid, the national centre for hygiene and safety), known as the Veilige Vijf (the Safe Five) of 2019. And it says what it is for: a basis for the protocols that organisations write themselves.

A district nurse asks: when do I apply hand hygiene at a client’s home?

A search on similarity finds at least four passages that resemble the question:

  • the five moments of hand hygiene from the SRI guideline;
  • the four moments from the RIVM guideline: room in, room out, before clean, after dirty;
  • the moments to wash your hands: after using the toilet, before and after (preparing) food, after contact with pets;
  • the Safe Five, if they still sit in a folder somewhere, on the same basic measures.

They all resemble the question. They do not all hold, and not all of them hold here. The grammar makes the difference visible. Below is how I would draw the links; an expert is welcome to shoot them down.

Origin. The RIVM guideline specialises the SRI guideline for district nursing. Five moments become four: at home, some moments do not apply or fall together. The whole home is the client’s environment, so “after contact with the environment” there usually means the moment you leave the home. For a question from district nursing the specialised version comes first, even though the general one resembles the question just as closely.

Basis. In the section on the moments it says: “Based on: SRI guideline Hand hygiene & personal hygiene of staff in long-term care (2023).” If that SRI guideline proved wrong, the recommendation would fall. That is “rests on”. The guideline also cites a study that translated the five moments into four. If that study proved wrong, the recommendation would not simply fall, but it would weaken. That is “informs”.

Structure. The moments to wash your hands are not part of the care tasks, but they complement them. The guideline says so itself: besides the moments in the SRI guideline there are moments “not directly connected with care tasks”. A complete answer names both.

Origin, once more. If a district nursing organisation writes its hygiene protocol on the basis of this guideline, that protocol derives from it; that is what the guideline is meant for. By rule 3, the protocol then rests on the SRI guideline as well, even if it never names it. When the SRI revises its guideline, the grammar can say which documents must be read again: everything that rests on it, directly or through a derivation.

Practice. A pocket card with the four moments puts the protocol into practice. It adds nothing to what is true. It makes it usable on the doorstep.

And the Safe Five? Replacing is not among the twelve. It is not a link between two sources that both hold. It says that one of them no longer holds, and that belongs to the version of a source. A search system that looks only at similarity finds the Safe Five as easily as when they were still in force.

Why care needs this

The passage that most resembles the question need not be the one that holds. In many fields that is an annoyance. In care, an answer that sounds like the right one but is the wrong one is worse than no answer. Anyone who asks a question at a client’s home must be able to build on the answer without someone standing beside them to say: that advice has been replaced, or: that is the rule for a ward, not for a home. The difference between five moments and four is not a detail. It is exactly the work the guideline did.

Knowledge in care comes in layers: a national guideline, a specialisation for a sector, the organisation’s protocol, the team’s work agreement. Search on words alone and you mix the layers up. Know the links and you know which layer speaks in this setting, and what it rests on.

The grammar does not decide which source leads. The organisation does. The grammar makes it known what rests on what, so that the choice has something to stand on.

The grammar belongs to no one field. It holds for a hygiene guideline as much as for a textbook. That is its strength and its limit. It says which kinds of link exist. It does not say which link lies between two actual documents. The people who know the field know that.

That is why this is a working document, and why I asked for it to be shot down without mercy. I set out claims as if they were true, so that someone can point to where they are not. A graph I fill on my own is a graph full of the assumptions of someone without the domain knowledge. So the division of labour is this: the people who know the field draw the links and test them; I build what records them and searches along them.

A wrong link cannot be seen in an answer. The answer looks just as good. That is why the grammar has to come under fire first, before Inora searches along it.

Quotations from my messages and from the guideline are translated from Dutch.

Do you work in care and use Inora? Sign-in and help go through your own organisation: your team’s project lead and ambassadors can help you.

Talk through what your team needs next.