La Trahison des mots

What’s the usual translation for embedding, in this particular context, in the French language? I mean, if we were to translate the concept as it appeared in A Neural Probabilistic Language Model, representing the concept back in 2003 and onwards, what is the usual practice in the Francophonie?

Recherche du terme français adéquat pour “word embedding”.

The most common established translation is plongement, usually in the phrases plongement lexical or plongement de mots (word embedding). It fits especially well because plongement is already the standard French term for embedding in mathematics, so the translation keeps the same link to the older mathematical sense that the English word has.

A few other practices are worth knowing:

Représentation vectorielle (de mots), or more loosely vecteur de mots, is a common descriptive alternative. It’s often preferred in teaching or general-audience writing, where plongement can sound opaque.
Représentation distribuée matches Bengio’s own framing in 2003. He described “distributed feature vectors” rather than using the word embedding, so a faithful rendering of that paper’s terminology would be something like vecteurs de caractéristiques distribués. That’s fitting, since Bengio is a Montreal-based francophone.
“Embedding”, left in English, is very common in practice. Researchers and engineers in France, Québec, and elsewhere in the Francophonie often use the English term in speech and informal writing, just as they say deep learning rather than apprentissage profond.

In a careful or formal text, plongement lexical is the safest choice, with représentation vectorielle as a clearer, more descriptive option.

Oh! Thank you for this:

Représentation distribuée matches Bengio's own framing in 2003. He
described "distributed feature vectors" rather than using the word
embedding, so a faithful rendering of that paper's terminology
would be something like vecteurs de caractéristiques distribués.
That's fitting, since Bengio is a Montreal-based francophone.

I’m here searching for a link with Magritte, and you told me already that the father of the embedding trick, representing tokens as vectors in multidimensional spaces, is a Montreal-based Canadian. =)

Untangling a playful link between two Belgian francophone figures.

There’s a nice link there, and it goes deeper than both men being francophone. Magritte was Belgian, and Bengio was born in Paris and built his career in Montreal.

The stronger connection is conceptual. La Trahison des images is about the gap between a representation and the thing it represents. The painted pipe isn’t a pipe, just marks on canvas that stand for one. An embedding is the same kind of object: the vector for “pipe” isn’t a pipe, and it isn’t even the word “pipe.” It’s a list of numbers that stands in for the word, and it’s useful only because of how it relates to the other vectors around it. You could caption any embedding with Ceci n’est pas un mot.

Magritte’s painting even carries a small warning that applies to language models. It’s easy to forget you’re looking at a representation and start treating it as the real thing. Much of the debate about whether models “understand” language is a version of Magritte’s question.

One small note to keep your framing accurate: Bengio’s 2003 paper is the usual starting point for learned embeddings in neural language models, but calling him the single “father” oversimplifies a bit. Hinton’s distributed representations came before it, and people like Collobert, Weston, and Mikolov shaped what came after. Still, a francophone at the origin of the idea makes a lovely pairing with Magritte.

Would this phrase be understandable in French: “La Trahison des mots”?

Untangling a French phrase’s double meaning.

Yes, it’s perfectly understandable and grammatical. It means “the betrayal of words” or “the treachery of words,” and most French speakers would immediately recognize the nod to Magritte’s La Trahison des images.

As in Magritte’s title, des is a little ambiguous. It most naturally reads as “the betrayal committed by words,” meaning words fail or deceive us. That’s the reading you’d want, and it’s the one the Magritte echo points to.

Depending on your angle, a couple of variations might fit your embedding idea even more closely:

La trahison des mots works if your point is that words themselves are already representations and not the things they name.
La trahison des vecteurs or La trahison des plongements works if your point is the second layer of representation: the model’s vectors stand in for words, which already stand in for things.

For a title, La trahison des mots is the most elegant of the three and reads well on its own.

Yes, for a catchy title, mots fits the bill. But I’d rather go with “La trahison des plongements” if it were an academic paper. Anyway, thank you for the valuable insights!

That makes sense. La trahison des plongements sounds more technical and would signal to researchers right away what the paper is about.

If you ever want both, academic titles often pair a catchy phrase with a precise subtitle, something like La trahison des mots : représentation et sens dans les plongements lexicaux.

You’re welcome, and good luck with the project!