Voynich Keywords Without Meanings: Local Vocabulary and Co-occurrence

Can a word be important before you know what it means?

Yes.

That answer sounds strange until you stop thinking of a keyword as a dictionary entry.

Imagine an unreadable text in which one token appears almost everywhere.

Another appears only in a cluster of plant pages.

A third appears mostly beside star-marked records.

A fourth repeatedly occurs in the same local regions as the second.

You still do not know whether any of them means “root”, “take”, “water”, “star” or something entirely different.

But their distributions are not equally informative.

The token that appears everywhere tells you little about where you are in the book.

The token that appears almost only in one local domain tells you much more.

That is the idea behind information-theoretic keyword analysis.

In 2013, Marcelo Montemurro and Damián Zanette applied such methods to a Voynich transcription and found that some tokens have strongly non-uniform distributions across the manuscript.

They also built networks from tokens that tend to occur in similar parts of the text and found affinities between visually defined sections based on shared word distributions.

The results are important.

They are also easy to overstate.

A statistically informative token is not automatically a semantic keyword in the ordinary sense.

Co-occurrence is not translation.

A cluster is not a topic name.

A network edge is not proof that two tokens mean related things.

Distribution can tell us where a token belongs before it can tell us what the token means.

That is enough to make distribution one of the most useful bridges between unreadable text and document structure.

Quick Read

One-sentence answer: some Voynich tokens cluster strongly in particular local regions and visual sections, while groups of tokens share similar occurrence profiles, showing that the manuscript has large-scale lexical organisation compatible with changing contextual domains; however, these “keywords” and co-occurrence networks are distributional signals, not decoded meanings, and their interpretation must control for page order, Currier regime, word families, document role and representation.

A token can be statistically informative because its occurrence is concentrated rather than uniform.
Montemurro and Zanette’s 2013 PLOS ONE analysis identified Voynich tokens whose distributions carry substantial information about local domains and visually defined sections.
Their method does not require knowing the underlying language first.
Some informative token families occur mainly in one or two sections.
Tokens with similar distribution profiles can be connected into co-occurrence or affinity networks.
Many strongly connected tokens also show visible morphological similarity, linking word-family structure with distributional structure.
The 2013 analysis found strong textual affinity between the pharmaceutical and herbal sections, which also share plant imagery.
It also found an affinity between the recipe/starred and astrological groupings under its section definitions.
Those results support non-random large-scale organisation.
They do not assign dictionary meanings to individual tokens.
The study used an EVA transcription and reordered scattered folios into visually defined section groups for part of its analysis, so section-order assumptions matter.
Current binding order is not guaranteed original production or reading order.
Currier A/B can create strong vocabulary differences independent of topic.
Document role such as labels versus running text can also create vocabulary differences independent of topic.
Word-family similarity can make co-occurring tokens look semantically related even if a generative or orthographic mechanism contributes to both.
A strong topic model should predict independent visual or historical relations on unseen material.

The manuscript gives us context signals.

It does not yet give us topic names.

What Is a Keyword When Nothing Is Readable?

In ordinary language, we think of keywords semantically.

An article about volcanoes contains words such as:

magma;
eruption;
lava;
crater.

But suppose every word is replaced by an unknown symbol code.

You can still notice that some coded words cluster in the same pages.

One coded form appears heavily wherever volcano diagrams occur.

Another appears throughout the book.

Without translating either, you can infer that the first is more context-specific.

Information-theoretic keyword analysis formalises that intuition.

The “keyword” is therefore initially:

a token whose distribution carries unusually high information about textual location or domain.

Meaning is a later inference.

Uniform Tokens and Local Tokens Do Different Jobs

A token distributed uniformly across the manuscript does not tell you much about which section you are reading.

It may behave like:

a common grammatical element;
a universal notation marker;
a high-frequency code group;
a generated structural component.

A token concentrated in one local domain carries more location information.

It may behave like:

specialised vocabulary;
a topic term;
a page-family code;
a source-specific orthographic form;
a Currier-state feature.

The distribution gives us a functional contrast.

It does not yet choose the cause.

The 2013 Result Was About Long-Range Organisation

Montemurro and Zanette’s central contribution was not a proposed translation.

It was showing that Voynich token distributions contain structured information across long spans of text.

Some tokens cluster in local domains rather than following approximately uniform occurrence.

This behaviour resembles what happens in meaningful texts when subject matter changes.

The authors compared the information profile with other sequences and argued that Voynich exhibits complex organisation compatible with genuine message-bearing language.

The important word is compatible.

A structured encoded text can show the same broad effect.

A highly organised generative system might too.

The result raises the burden on simple random-text explanations.

It does not by itself decode semantic content.

Local Vocabulary Is Not the Same as Local Word Shape

This batch already gave us a word-family article.

That article studies whether nearby tokens are visually similar.

This article asks a different question.

Which token identities are concentrated in which regions?

A token can be locally concentrated without resembling its neighbours.

Another can resemble nearby tokens while appearing throughout the manuscript.

Form similarity and occurrence locality are related but distinct statistics.

This separation matters because a local-generation model may produce both.

A topical natural-language model may produce occurrence locality strongly but form similarity only according to morphology.

Comparing the two can discriminate mechanisms.

Co-occurrence Means “Used in Similar Places” Before It Means “Semantically Related”

If two tokens repeatedly appear in the same local regions, their occurrence profiles are similar.

In known languages, semantically related content words often behave this way.

doctor, patient, hospital and treatment cluster in medical passages.

Therefore distributional similarity can become evidence of semantic affinity.

But the inference has alternatives.

Two forms may be morphological variants of the same underlying unit.
Two code groups may belong to the same encoding table.
One scribe may favour both.
A local generation rule may create both from the same seed family.

So the safe statement is distributional:

these tokens inhabit similar textual regions.

The semantic interpretation needs another witness.

Morphological Similarity Inside the Co-occurrence Networks Is Fascinating

Montemurro and Zanette noted that strongly connected informative tokens often also resemble one another in form.

Some share apparent prefixes.

Some share endings.

This could be exactly what natural-language morphology predicts.

Related words occur in related contexts.

It could also be exactly what a local-generation model predicts.

Related-looking tokens are generated near one another, so their distributions overlap.

This is a beautiful example of observational equivalence.

The same network pattern can support two different causal stories.

Progress requires a feature on which the stories diverge.

The Herbal–Pharmaceutical Link Is One of the Most Intuitive Section Results

In the 2013 section-affinity analysis, the strongest textual link connected the pharmaceutical and herbal groups.

That is interesting because the imagery independently shares plant material.

The herbal pages show whole or near-whole plants.

The pharmaceutical pages show detached plant components and vessels.

Textual affinity and visual affinity therefore point in a compatible direction.

This is stronger than either resemblance alone.

But it still does not prove a literal pipeline from plant identification to preparation.

The sections can share vocabulary because they discuss related domains without one being an operational continuation of the other.

Affinity supports relationship.

It does not identify relationship type.

The Recipe–Astrological Link Is More Interesting Because It Is Less Obvious

The 2013 analysis also reported a strong relationship between the recipe/starred and astrological groupings under its section definitions.

This is useful because the visual relationship is not simply “both show plants everywhere”.

Star-like motifs appear across the relevant visual environments, but the sections otherwise differ substantially.

A real semantic relationship is possible.

Timing?

Classification?

Shared records?

Cross-reference?

But the result must remain distributional until one of those semantic bridges is independently demonstrated.

Our earlier work did not support promoting zodiac or stars into a proven universal control layer.

The section affinity keeps a relationship hypothesis alive without resurrecting the stronger failed claim.

Section Definitions Are Convenient but Not Neutral

The conventional herbal, astrological, biological, pharmaceutical and recipe labels are based mainly on visual appearance.

They are useful navigation categories.

They are not proven medieval chapter titles.

Therefore a statistical method that measures how well words identify these sections partly tests whether textual distribution aligns with modern visual grouping.

That is a valuable test.

It should not be restated as:

the manuscript has five proven semantic chapters.

Modern section names are hypotheses about visual organisation.

Textual alignment strengthens their relevance.

It does not make their names literal.

The 2013 Analysis Reordered Scattered Folios for Part of the Section Study

This methodological detail matters.

The study grouped scattered folios with the visual section to which they were assigned rather than leaving every page in current binding order for the section-based analysis.

That is a reasonable analytical choice when asking about section vocabulary.

It also means the result is conditional on the section assignment.

This is another reason not to turn distributional networks into original reading-order claims.

The manuscript’s present order, production order, original binding order and conceptual order remain separable problems.

Currier A/B Is a Major Confounder for Topic Inference

Suppose one token appears mainly in Currier B.

Currier B is concentrated in particular manuscript regions.

The token now looks section-specific.

Is it a topic word?

Or a B-regime orthographic form?

Or a grammatical element more common in one regime?

This is why topic analysis should control for Currier where possible.

The most useful candidate keywords may be forms that remain section-specific even after regime differences are accounted for.

Again, crossing cases are more informative than perfect correlations.

Scribal Hand Can Also Masquerade as Topic

If one proposed scribe works mainly on one section, personal spelling habits become section-specific statistics.

A token family may look like specialised vocabulary when it partly reflects one writer’s orthography.

The exact scribe count remains debated, which makes this control even more important.

Topic, Currier regime and hand should therefore be mapped separately before one token is given a semantic role.

The strongest keyword candidate is one whose local concentration cannot be reduced easily to writing style alone.

Document Role Can Masquerade as Topic Too

The label article adds another confounder.

Suppose a section contains many labels.

Labels have a different vocabulary distribution from running text.

A section-specific vocabulary difference may therefore arise partly because the section contains a different proportion of labels.

Likewise, Quire 20’s short star-marked records may have a register unlike long paragraphs.

A true topic model should distinguish:

what the text is about

from:

what kind of text unit it is.

Local Generation Can Mimic Topic Bursts

Now bring in the generation hypothesis.

If a writer creates new tokens by modifying nearby ones, local vocabulary bursts appear naturally.

A family of related forms can dominate one region and then fade as the generator drifts to another family.

This can resemble semantic topic change.

How do we separate them?

A semantic-topic model predicts relationships to independent content.

Visual section.

Repeated depicted object.

Genre.

A pure local-generation model predicts stronger dependence on textual neighbourhood and edit distance.

The two may coexist.

The discriminator is whether distribution predicts something outside the token stream.

A Cipher Can Preserve Topic Structure

Meaningful plaintext about changing topics can remain distributionally organised after encryption.

Whether the structure survives visibly depends on the cipher.

A simple substitution preserves token identity strongly.

A homophonic or verbose system can blur it while still leaving local state or code-family patterns.

Therefore topic-like clustering does not distinguish plaintext from ciphertext by itself.

A serious encoding model should demonstrate how the proposed cipher transforms known topic-bearing plaintext and whether comparable long-range distributions survive.

The 2025 Naibbe work makes constructive cipher comparisons increasingly important across the entire Voynich research programme.

A Statistical Keyword Is Not Necessarily a Content Word

Suppose one Currier-B suffix occurs almost only in B-heavy sections.

Its distribution carries information about section location.

It may therefore rank as statistically informative.

But semantically it could be a grammatical ending, not a topic word.

This illustrates why “keyword” can be misleading to ordinary readers.

The method identifies location-informative forms.

Some may eventually turn out to be content-bearing vocabulary.

Others may encode grammar, scribal practice or regime.

Semantic interpretation must be earned separately.

Co-occurrence Networks Are Maps of Shared Context, Not Dictionaries

A network is visually persuasive.

Nodes.

Edges.

Clusters.

Human eyes immediately see concepts.

This cluster must mean plants.

That cluster must mean astronomy.

The graph itself does not contain those labels.

It contains relationships generated from occurrence patterns.

A responsible interpretation therefore has two layers:

These token forms share contextual distribution.
A semantic hypothesis may explain why.

Skipping the first layer turns statistical geometry into invented vocabulary.

The Most Powerful Future Use Is Prediction Across Modalities

Suppose one distributional cluster is discovered without looking at illustrations.

Its tokens concentrate on a subset of pages.

Only afterward do researchers inspect those pages and discover a repeated visual feature that had not been part of the section classification.

That is strong.

Text has predicted image structure.

Or a network predicts that one pharmaceutical label family should correspond to a whole-plant family elsewhere.

The visual match is then found independently.

Now distribution becomes a semantic bridge.

This is much stronger than naming a cluster after the images used to create it.

Local Vocabulary Can Help Reconstruct Missing Context

The manuscript is incomplete.

Missing leaves interrupt several regions.

If surviving pages on both sides of a gap share an unusually specific vocabulary profile, that may suggest contextual continuity.

If profiles differ sharply, perhaps the lost material lay near a real transition.

This cannot reconstruct the missing words.

It can constrain the type of textual environment around the loss.

Distribution is therefore potentially useful for codicology as well as semantics.

Again, current order uncertainty must remain visible.

Local Vocabulary Can Help Test Page Reordering

Suppose a folio is suspected to be misplaced in the current binding.

Its vocabulary profile can be compared with candidate neighbouring regions.

If it strongly resembles one distant physical or visual group, that becomes evidence of affinity.

But affinity is not original adjacency.

Two chapters on the same topic can be distant.

Two source packets can share vocabulary.

Therefore distribution can nominate candidate relationships.

Physical bifolium and quire evidence must still decide what kinds of rearrangement are possible.

Topic Drift May Be More Continuous Than Modern Sections Suggest

A book does not need to change subject at one sharp boundary.

Vocabulary can drift gradually.

One module can bridge plant and pharmaceutical content.

A zodiac discussion can blend into health timing.

A record section can summarise earlier material.

Therefore the most realistic topic map may be continuous or overlapping rather than five hard semantic boxes.

This fits the broader conclusion from retained research:

the manuscript is better modelled by overlapping axes than by one master segmentation.

Visual section, Currier state, hand, document role and vocabulary domain can overlap without being identical.

The Strongest Keyword Is the One That Predicts Something Else

A token is interesting if it is section-specific.

It becomes more interesting if its family has a stable morphological pattern.

More interesting again if it recurs in labels.

And far more interesting if its occurrence predicts an independently defined visual class.

This creates an evidence ladder:

local concentration;
stable co-occurrence;
morphological family;
cross-register recurrence;
independent visual or historical prediction;
deciphered semantic value under fixed rules.

The ladder protects distributional research from claiming meaning too early while still letting it make progress toward meaning.

What Survives the Distributional Work

Voynich token usage is strongly non-uniform across the manuscript.
Some tokens are much more informative about textual location than others.
Local contextual domains can be detected without decipherment.
Groups of informative tokens share similar occurrence profiles.
Many connected tokens also share visible morphological features.
Textual section affinities can align with visual affinities, especially herbal and pharmaceutical material.
The manuscript therefore has real large-scale organisation beyond simple character-level regularity.
Current visual section definitions, Currier regime, hand and document role can confound semantic topic inference.
Local-generation and encoding models can reproduce some forms of distributional organisation and remain viable alternatives.
Distribution becomes much stronger semantic evidence when it predicts independent visual, historical or functional relations.

What Does Not Survive as Established Knowledge

The top statistical keywords have decoded meanings.
A co-occurrence network is a semantic dictionary.
Every cluster corresponds to one topic.
The conventional five visual sections are proven medieval semantic chapters.
The herbal–pharmaceutical affinity proves a literal plant-to-remedy pipeline.
The recipe–astrological affinity proves zodiac control over recipes.
Local vocabulary alone proves natural language.
Local vocabulary alone rules out generated or encoded text.
Current binding order is original conceptual order.

The organisation survives.

The topic names do not.

A Better Distributional Analysis

Specify transcription and tokenisation.
Separate location-informative tokens from semantic claims.
Measure locality at several scales rather than one section size.
Control for Currier A/B.
Control for proposed scribal hand where possible.
Separate labels, prose, circular text and star-marked records.
Keep visual section definitions provisional.
Compare occurrence locality with word-family edit distance.
Use physical codicology when testing page order.
Demand independent cross-modal predictions before assigning meanings.

What Would Count as a Real Keyword Breakthrough?

Imagine an unsupervised distributional analysis identifies a token family concentrated on pages that do not share an obvious modern section label.

Only afterward, blind visual comparison reveals that those pages share a specific root morphology or vessel class.

The text has predicted a previously unrecognised visual relation.

Now suppose a decipherment independently assigns the token family a meaning consistent with that class and succeeds on withheld pages.

That would be genuine convergence.

Or imagine a proposed Currier transformation removes what looked like section keywords, proving they were orthographic regime markers instead.

That too would be a breakthrough because one hidden cause has been separated from another.

The point is not to make every keyword semantic.

It is to discover what variable each informative token is actually informative about.

Primary School: Find the Secret Topic Without Reading the Words

Replace every noun in three short passages with invented symbols.

One passage is about cooking.

One about football.

One about plants.

Keep the same invented symbol each time the same noun appears.

Ask the child which symbols are local to one passage.

They can discover topic markers without knowing the hidden vocabulary.

That is distributional inference in miniature.

Lower Secondary: Context Without Translation

Give students a table showing where ten invented tokens occur across five chapters.

Ask which token tells you most about which chapter you are in.

A token appearing everywhere is weak.

A token appearing only in Chapter 4 is strong.

Then ask:

Do we now know what the Chapter-4 token means?

No.

We know where it belongs.

The distinction is the heart of this article.

Upper Secondary: Build a Co-occurrence Network

Give students a small coded corpus.

Connect two tokens if they repeatedly occur in the same passages.

Clusters appear.

Now ask students for two explanations.

The cluster contains semantically related words.
The cluster contains forms generated by one local copying rule.

What new data would separate them?

The learner discovers why a network is evidence of relation without automatically being evidence of meaning.

JC and Adult Readers: Think in Conditional Mutual Information

At a higher level, the central question is not whether a token predicts section.

It is whether the token still predicts section after other variables are known.

Given Currier regime, does it still add location information?

Given scribal hand?

Given label versus prose role?

Given word-family membership?

The residual information is closer to a genuine topic signal.

This is the mature version of keyword analysis:

not “where does this token occur?”, but “what hidden variable best explains why it occurs there?”

A Parent and Teacher Guide

Find concentrated tokens before naming them.
Separate location information from meaning.
Compare co-occurrence with visual context.
Control for writing style and document role.
Remember that morphological relatives can cluster mechanically.
Use section labels as provisional navigation categories.
Reward predictions that discover new cross-modal relationships.

The wider lesson is invaluable:

you can learn the shape of context before you know the names of the things inside it.

Reader Checklist: Before You Call a Voynich Token a Keyword

Which transcription and tokenisation are used?
At what spatial scale is the token concentrated?
Is the “section” visually defined or independently textual?
Does Currier regime explain the concentration?
Does proposed hand explain it?
Does label versus prose role explain it?
Is the token part of a dense word family?
Could local generation create the same burst?
Could an encoding state create it?
Does the token co-occur with independent visual features?
Does the relationship survive on pages not used to define the cluster?
Are you assigning a meaning that the network itself never supplied?

Frequently Asked Questions

Can the Voynich Manuscript have keywords if we cannot read it?

It can have statistically informative tokens whose distributions are concentrated in particular textual domains. Calling them semantic keywords is a later interpretive step.

What did Montemurro and Zanette find?

Their 2013 information-theoretic analysis found complex long-range token organisation, identified highly location-informative forms and built co-occurrence networks and section-affinity relationships from their distribution profiles.

Did they decipher the keywords?

No. The analysis identifies distributional importance and candidate semantic relationships without assigning accepted lexical meanings.

Why is the herbal–pharmaceutical connection important?

Because those sections are strongly related by token distributions and independently share plant imagery. The convergence supports a relationship while leaving its exact function unresolved.

Does local vocabulary prove topics?

No. Topic is one explanation. Currier regime, hand, document role, encoding state and local generation can also create concentrated token distributions.

Does co-occurrence prove two words have related meanings?

No. It shows that they occur in similar contextual regions. Semantic affinity is plausible but requires additional evidence.

Why does page order matter?

Locality and long-range distribution depend on sequence. Current binding order is not guaranteed original production or reading order, and some analyses group scattered folios by visual section for specific questions.

What is the strongest current conclusion?

Voynich vocabulary is organised across multiple scales. Some token forms are strongly specific to local domains, and these distributional structures align partly with visual organisation. Their meanings remain undeciphered.

Research and Further Reading

The Final Idea

The Voynich Manuscript can tell us what changes together before it tells us what anything is called.

One token lives mostly here.

Another follows it.

A family blooms in one section and disappears in another.

Two visually related page groups share vocabulary.

Those are not meanings.

They are constraints around meaning.

And constraints are exactly what an undeciphered manuscript needs.

The final solution will not begin from a blank page.

It will inherit a map of where forms belong, which forms travel together, which registers differ, which visual sections share vocabulary and which correlations survive alternative explanations.

A keyword without a meaning is not useless. It is a coordinate waiting for a language.

Continue Through the Voynich Research Map

This article is one specialist node in eduKateSG’s larger Voynich research library. Return to the canonical master to see the physical, visual, textual and historical evidence in one continuous argument.

Structure is evidence; structure is not translation. The master keeps resemblance, statistical structure, historical possibility and decipherment claims at separate evidentiary levels.

Voynich | Everything eduKate Knows and Tested | Keywords Without Meanings: Local Vocabulary, Co-occurrence and Topic Signals