Eleven lenses on a single word

In a brief 1942 essay on John Wilkins' analytical language, Jorge Luis Borges describes a 'certain Chinese encyclopaedia' that divides animals into categories such as 'those that belong to the emperor', 'embalmed ones', 'those drawn with a fine camel-hair brush', and 'those that from a long way off look like flies'.

The list is fictional, but its scandal is real: every taxonomy looks self-evident from inside its own culture, and absurd from outside. Borges teaches us to read classification as poetry — and to suspect that our own categories would look just as strange to a future reader.

Michel Foucault argued that knowledge and power do not stand opposite each other: institutions produce knowledge, and that knowledge in turn enables new forms of power. The clinic, the school, the prison, and the census do not merely describe people — they produce the categories (the patient, the pupil, the delinquent, the population) through which people become governable.

On this view, classification is never neutral. Every category opens some interventions and forecloses others. Asking 'what is this?' is inseparable from asking 'who decided, and what do they gain by deciding so?'

Structuralism, especially after Saussure, holds that signs do not mean by pointing at things, but by differing from other signs in the same system. 'Day' means what it means partly because 'night' exists. Removed from its system, a word loses its hold on us.

This is a useful counterweight to naive realism about categories. It reminds us that the value of a term is structural before it is referential.

Post-structuralist thinkers, from Derrida to Butler, push further: the relation between sign and meaning is not stable but deferred, contextual, and political. Identities and categories are performed and re-performed; their boundaries leak.

For an ontology engineer, this is not nihilism. It is a practical warning that any frozen schema captures one moment of a moving usage.

Modern AI systems inherit the categories present in their training data, which are themselves the residue of institutions, languages, and decisions about what was worth recording. When a model classifies a face, ranks a job application, or labels a sentence as 'toxic', it is operationalising a choice made elsewhere.

The task is not to imagine a category-free system, but to make the categories explicit, contestable, and historically aware.

A retrieval system encodes assumptions twice: in what it indexes and in how it ranks. 'Relevance' is not given by the world; it is constructed by a chain of choices about tokenisation, embedding, similarity, and re-ranking. RAG systems amplify this by feeding retrieved fragments back into generation.

This is why the ontology behind a retrieval system matters as much as the model on top of it.

Embeddings represent terms as vectors in a high-dimensional space, where 'closeness' captures co-occurrence rather than membership in a defined set. This permits graceful handling of ambiguity and analogy, but at the price of opacity: the categories are real, but no longer named.

Embeddings do not abolish ontology; they hide it. Bringing it back into view is part of what this app demonstrates.

Words move. 'Hacker' once meant a playful explorer of systems; later, a criminal; today, often a security professional. 'Computer' once meant a person doing arithmetic. Categories drift as institutions, technologies, and politics shift around them.

Any honest ontology must record this drift, not erase it.

Schools, clinics, militaries, and statistical bureaus measure people against a norm and then act on the deviation. The normal curve is not just descriptive — it is prescriptive, attached to interventions and incentives.

Foucault called this 'normalisation'. It is one of the most powerful and least visible operations of modern classification.

A choice of categories decides who is counted, what is funded, what is criminalised, and what is invisible. There is no apolitical ontology. There are only ontologies whose politics are explicit and ontologies whose politics are hidden.

This app is a small, deliberate attempt at the former.