The Meanings Our Machines Place Together

How does a machine decide that two expressions in different languages mean the same thing?

And what disappears when it does?

I arrived at those questions through a comment beneath a video from Dr. Tressie McMillan Cottom.

She had asked whether there was a word for a particular kind of help: the kind offered by someone who sees your difficulty but not the solution you have already constructed. The kind that interrupts your competence and creates the very problem it intended to prevent.

People understood the experience immediately.

Then they began offering language for it.

One commenter offered the Yoruba expression atọ́rọ̀ṣe bọ̀rọ̀jẹ́.

That comment stopped me. Not because I had discovered the definitive word. I had no basis for making that claim. I did not yet know what the expression meant to Yoruba speakers, how it was used, or whether it described the helper, the intervention, or the resulting mess.

What caught my attention was the possibility that another language had organized the experience differently.

English speakers could recognize what happened. We could explain it. We could judge it. But we seemed to need several expressions to hold all its parts.

That possibility became the center of my last post, The Words Our Machines Cannot Find.

But the encounter also raised a different kind of question.

Multilingual AI systems are designed to recognize related meanings across languages, even when the words themselves look nothing alike. To do that, they must decide which differences can be abstracted away and which ones still matter.

So what happens when a machine places two expressions close together because their meanings appear similar?

What should remain distinct?

From Matching Words to Matching Meaning

Search systems have historically relied heavily on lexical signals: the words someone enters, the words contained in a document, and the statistical relationships between them.

That works well when the query and the relevant material use similar language.

But people do not always search with the right words.

We describe what we remember. We use an approximation. We ask a question instead of naming the subject. Sometimes we know the meaning we are reaching for but not the word that would retrieve it.

Crossing languages makes the problem harder. The relevant result may contain none of the words in the original query. It may use a different writing system, grammatical structure, or way of dividing the concept itself.

Semantic search offers another way to find a match.

Models can convert queries and documents into numerical representations called embeddings. Texts the model has learned to treat as semantically related are placed closer together in a multidimensional space.

The system can then retrieve something because it appears to concern the same subject, even when it does not contain the same words.

Dense retrieval systems demonstrated that learned vector representations could retrieve relevant passages without relying entirely on traditional keyword matching. [See: Dense Passage Retrieval for Open-Domain Question Answering]

Retrieval-augmented generation raised the stakes. In these systems, retrieval does not merely determine which links appear on a results page. It helps determine which information a generative model receives before producing an answer. [See: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks]

Semantic retrieval does more than help a system find information.

It helps determine which information the system considers relevant enough to use.

Meaning Isn’t Always a 1-to-1 Swap

In multilingual systems, the ambition is greater still.

A system must recognize semantic relationships between texts that may share neither vocabulary nor structure. An English description and a Yoruba expression can look entirely different while pointing toward related territory.

The connection is semantic, not typographic.

Multilingual embedding models attempt to capture that connection by placing translated or semantically corresponding texts near one another. This can allow someone to search in one language and retrieve material written in another. It can help translation systems identify corresponding sentences and allow capabilities learned from languages with abundant training data to transfer to languages with less.

Researchers demonstrated one version of this approach with Language-agnostic BERT Sentence Embedding, or LaBSE. The model was designed to place translated or semantically corresponding sentences near one another, supporting cross-lingual retrieval across more than one hundred languages. [See: Language-agnostic BERT Sentence Embedding]

That is an extraordinary technical achievement.

It is also an act of abstraction.

The system becomes useful across languages because it learns to look past many of the features that distinguish one linguistic form from another.

But proximity is not equivalence.

Two expressions can occupy the same semantic neighborhood without living at the same address.

One may locate the meaning in the person who intervened. Another may emphasize the act itself or the condition left behind. One may carry humor. Another may carry judgment. Either may change depending on who says it, to whom, and under what circumstances.

The trouble begins when similarity is presented as substitution.

A system may produce a concise English equivalent that preserves the broad subject while losing what the original expression asks its speakers to notice. A translation can be defensible at the level of topic while remaining incomplete at the level of human meaning.

That does not make the translation useless.

It means the limits of the translation belong in the answer.

A Shared Space Is Still a Point of View

The phrase “language-agnostic” can suggest that a model has reached some neutral layer of meaning beneath language.

I understand the appeal of that idea. If language is the variable, perhaps meaning is the constant underneath it.

The technical reality is more complicated.

Researchers examining multilingual models have found that their internal representations can support cross-lingual transfer while still retaining information about the language that produced the text. One study of multilingual BERT and XLM-R found that contextual representations were only moderately language-neutral by default. [See: On the Language Neutrality of Pre-trained Multilingual Representations]

A model does not simply remove language and reveal pure meaning underneath.

It constructs a representation shaped by its architecture, training data, optimization objectives, and evaluation tasks.

A shared semantic space is not a naturally occurring map of human meaning.

It is a map built for a purpose. Its structure reflects choices.

Which texts appeared in the training data? Which languages had enough examples to establish stable patterns? Which translations were used to align them? Which variations were treated as equivalent, and which were treated as errors?

A model may technically support many languages while supporting them unevenly. Coverage tells us that a language was included. It does not tell us how well the system represents regional usage, code-switching, tonal distinctions, humor, oral conventions, or culturally specific concepts.

Meaning can also change before the model receives it.

When I first tried to investigate atọ́rọ̀ṣe bọ̀rọ̀jẹ́, I entered it without the tonal marks and underdots used in written Yoruba.

I searched for “atorose boroje.”

Almost nothing came back.

When I returned to the original comment and copied the expression as it had been written, I could see what my transcription had erased.

The marks were not decoration.

They were part of the language.

By the time a multilingual system produces an answer, a speaker has chosen an expression, someone has written or transcribed it, a keyboard has made certain characters easier to enter than others, a tokenizer has divided the text into units, and a retrieval model has ranked the possible matches.

Each layer creates another opportunity for meaning to be preserved, transformed, or lost.

Uncertainty Belongs in the Product

A multilingual system will not always know whether two expressions are equivalent.

Good.

Trust is not built by answering every question with the same polished confidence. It is built when a system helps the user see where its knowledge ends.

Admitting uncertainty gives the user something useful: a clearer basis for deciding how much confidence to place in the answer.

If several expressions are related but carry different implications, the system can explain those distinctions instead of collapsing them into one confident translation.

If an expression is poorly documented, it can say so.

If a spelling appears incomplete or ambiguous, the interface can offer possible forms instead of silently selecting one.

If diacritics materially affect meaning or retrieval, the system can help users restore them.

If a translation captures the general sense while potentially losing humor, register, or cultural context, that limitation can accompany the answer.

None of this requires the system to refuse every imperfect request.

Translation has always involved approximation. People communicate every day across partial knowledge, uncertain equivalence, and imperfect vocabulary.

The problem is not approximation.

The problem is approximation presented as certainty.

More data alone cannot decide which distinctions matter within a community or when an approximation has changed the meaning. Those judgments require linguistic and cultural knowledge.

They also require the participation of speakers who can identify where a translation is technically plausible but socially wrong, where a spelling variation is recoverable, and where two apparently similar expressions should remain distinct.

Speakers cannot be treated only as sources of training data.

They are also experts in the boundaries the system is attempting to model.

Admitting those boundaries does not make a system less capable.

It makes its capabilities easier to trust.

What Should Survive the Journey?

And this, friends, is where I geek out.

There is a promise in multilingual AI that I find genuinely exciting, in part because it remains technically unsettled.

A system might do more than help us understand words written in another language. It might help us encounter distinctions that another language has learned to preserve.

That is a much harder task.

It requires a system to recognize what two expressions share without assuming they are interchangeable.

Another language may gather ideas that English keeps separate. It may locate the significance of an experience somewhere we did not expect. Encountering that difference can change what we notice, even if we never adopt the expression ourselves.

This is what cross-lingual discovery can offer at its best: access to another language, certainly, but also an encounter with another way of organizing experience.

That second possibility depends on preserving enough difference for the other language to remain instructive.

If every expression is reduced to its nearest familiar approximation, a system may improve access while narrowing discovery.

If similarity becomes sameness, translation stops being an encounter with another way of making meaning. It becomes a process of making everything familiar.

The technical challenge is not simply to create a shared semantic space.

It is to decide what must remain distinct within it.

A capable multilingual system should recognize when two expressions are related.

A trustworthy one should also help us understand why they are not interchangeable.

Meaning needs to travel.

But the differences that give it shape need to arrive too.