The Pivot

TAM-INS.06 · The Insufficient · The Approximate Mind

Nasrin Zargar has interpreted in immigration court in Newark for nineteen years. She keeps a spiral notebook on the table beside the microphone, and in the right margin of each page she writes short notes to herself that never enter the record: he softened that, she is addressing the judge and not me, this is a form he would only use to a stranger. The transcript gets the sentences. The margin gets everything the sentences were doing.

On a Tuesday in March she interprets for Sohrab Nazari, thirty-four, who is describing what happened to his brother. He speaks in Dari, and twice he uses the form that marks a thing he did not see himself but was told. In the transcript this becomes “my brother was taken.” Nasrin writes in the margin: he is saying he did not witness it. Nobody reads the margin. The adjudicator will later note that the account is inconsistent, because in an earlier interview the applicant appeared to claim direct knowledge.

The court has an interpreter, which is what the statute requires. What the statute does not require, because it has no words for it, is a record of the relations the sentences carried.

The Triangle and Its Apex

#

The four-hundred-year dream of a notation in which meaning is unambiguous and reasoning becomes calculation belongs to Pax Interlingua, elsewhere in this collection, which spends it properly. What matters here begins in 1968, when Bernard Vauquois drew a triangle.

The base of the triangle is direct translation, word against word. The middle is transfer, where a sentence is parsed into the structure of the source language and then rewritten into the structure of the target. The apex is interlingua: analyze the source all the way up into a representation that belongs to no language at all, then generate the target from it. At the apex, translation stops being translation. It becomes encoding and decoding through a neutral middle.

Everyone understood the apex was the prize. Two languages need one pair of components rather than one translator; a hundred languages need a hundred pairs rather than nearly ten thousand. Decades of work went into specifying the middle. Case grammars, semantic primitives, conceptual dependency structures, formal ontologies. Committees argued about whether “give” decomposes into transfer of possession and whether transfer of possession decomposes further.

Every designed interlingua failed. Not partially. The projects ran, produced systems, demonstrated on constrained domains, and did not generalize, and by the late nineties the field had largely stopped trying and moved to statistical methods that ignored meaning entirely.

Then a multilingual model was trained on many languages at once, and something that behaves like an interlingua appeared in the middle of it. Sentences with the same content in different languages land near each other in the same space. Translation between pairs the system was never trained on works. The apex of the triangle exists, in production, at scale, right now.

The dream was a notation in which meaning could be calculated, and it arrived without a specification.

Nobody wrote it. Nobody can read it. It fell out of the training and it is running under a substantial fraction of the sentences that cross a linguistic boundary on any given day, including the ones in Nasrin’s courtroom when the applicant is speaking a language for which no interpreter was available and the system fills in.

What the Philosophers See

#

To a philosopher of mind this is the first empirical instance of something that had only ever been argued about. The language of thought hypothesis proposes a representational medium underneath natural language in which thinking actually happens. It has been contested for fifty years by people who could not point at one. Here is a candidate that can be probed with a vector instead of a thought experiment.

Quine argued that translation is indeterminate: that multiple incompatible translation manuals can fit all possible behavioral evidence, and nothing in principle picks between them. A working pivot is a wager against that argument, and it is a wager being settled by engineering rather than by philosophy. Not answered. Settled, in the sense that the infrastructure now behaves as though the question has an answer, which is a different thing and a worse one.

Harnad’s symbol grounding problem asks how symbols in a formal system acquire meaning rather than merely relations to other symbols. The multimodal turn, where the same space holds images and text and audio, is a partial answer that nobody has finished assessing, because the systems arrived faster than the assessment.

None of this is remote from the courtroom. Quine’s indeterminacy is usually taught with a linguist in a field and a rabbit running past. The working version is an adjudicator in Newark deciding whether two accounts of the same event are inconsistent, where the manual that produced the second account is one of several that would fit, and where nothing in the record indicates that a manual was applied at all. The philosophical problem and the administrative one are the same problem at different volumes.

I wonder whether a discipline can be said to have lost an argument it was never told had been submitted for judgment.

What the Anthropologists See

#

A linguistic anthropologist asks a different question. Not whether the pivot is metaphysically sound. What a deployed one does.

An utterance says something and it also does something. It marks who is speaking, to whom, in what relation, with what claim to knowledge, at what social distance. Silverstein’s term for the second function is indexicality, and Hymes made the case that competence in a language is mostly competence in the second function rather than the first. A speaker who produces grammatical sentences and cannot mark deference is not a fluent speaker. They are a person who will be misread constantly.

The pivot preserves the first function very well. It handles the second function by dropping it.

This is not an implementation gap that a better model closes. The middle is a space where sentences with the same content sit near each other, and content is exactly the first function. Deference, hedging, evidential marking, the distance between an intimate register and a formal one: these are not content. They are relations between the speaker, the addressee, and the claim, and a representation organized around what a sentence is about has nowhere to put them.

Irvine and Gal give the precise name for what follows. Erasure is the process by which an ideology renders distinctions invisible, not by suppressing them but by having no place for them in the scheme. The distinctions do not appear as losses. They do not appear at all.

A layer that carries everything a sentence says and nothing it means about the speaker is a filter with an ideology, and the ideology is that the second function was never load-bearing.

Sohrab said he had not seen it. The transcript says his brother was taken. The margin says what was dropped, and the margin is in a notebook.

What the Typologists See

#

There is a careful literature adjacent to all of this, and it is disciplined in ways the surrounding discourse is not.

Semantic map research asks whether the meanings languages group together under a single word cluster in non-arbitrary ways across the world’s languages. Colexification research asks the same question at scale: which meanings do languages repeatedly bundle, and which do they repeatedly keep apart. The answer these fields return is partial structure. Real regularities, repeatedly found, nowhere near a universal conceptual inventory, and reported with the error bars attached.

That literature is the correct comparison class for any claim about what the pivot demonstrates. It is also the literature least likely to be cited when the claims are made, because it is slow, it qualifies everything, and its findings do not support the interesting version.

The Slide

#

Here is the inference that will be made constantly over the coming decade, by people with an interest in making it.

Models trained on overlapping corpora converge on a shared representation. Therefore there is a universal human conceptual substrate, and we have found it.

Refuse this. The convergence is evidence about corpora. Systems built from overlapping human output, using similar architectures, optimized against similar objectives, converge, and the convergence tells you about the output and the architectures. It does not tell you about minds. It is agreement among artifacts, and agreement among artifacts built from the same material by the same methods is the least surprising result available.

The slide is attractive because it converts an engineering artifact into a discovery about human nature, and discoveries about human nature justify a great deal. Specifically, they justify not asking what the middle drops. If the pivot has found the substrate that all human concepts share, then whatever failed to survive the crossing was never conceptual to begin with, and the loss is reclassified as noise before anyone has counted it.

That reclassification is where Sohrab’s evidential marking goes. Not deleted. Ruled out of scope, by an argument nobody in the hearing room made or heard, on the strength of a finding about the geometry of a vector space.

The Operation

#

The Insufficient does not correct systems. It equips them to report their own insufficiency, which is a narrower and more achievable thing.

The Traditions named seven operations and treated them as a closed set. They were closed in a particular way. Each was drawn from a tradition the development ecosystem was not built to see, and each was aimed at the same object: a category applied to a person. Reification, atomization, decontextualization, misframing. Seven ways of asking whether the box fits the life.

What follows is an eighth, and it is an eighth rather than a correction to the count because the object moves. It comes from the same place the others do, a tradition outside the ecosystem’s line of sight, and linguistic anthropology qualifies as squarely as Madhyamaka or Ubuntu. What changes is what it is aimed at. The seven interrogate a classification. This one interrogates a component, and the component is one that no classification passes through, because it sits underneath the sentence in which any classification would have to be stated. The set was closed against traditions. It was never closed against layers.

The indexical audit. Before a pivoted output is treated as equivalent to its source, the system enumerates the features of the source that mark relation rather than content: honorific and deference forms, evidential marking, hedges and mitigators, register level, address terms, the distance between what was said and what was meant by saying it. For each one it determines whether the target carries a corresponding feature. Where it does not, it emits a flag.

The flag does not repair anything. It does not guess what the feature was doing. It states that a relation-marking feature was present in the source, that the target has no place for it, and that the system cannot say what was lost. Its output is a list, not a correction.

This is executable. Every feature named is detectable by existing methods in the languages where it has been described, and the honest form of the operation includes reporting that for many languages it has not been described, which is itself a finding about who has been studied.

It would have put a line in the record of Sohrab’s hearing where Nasrin’s margin note is. It would not have told the adjudicator what the evidential form meant. It would have told him there was one, and that the English does not have it, and that the inconsistency he is about to note may be an artifact of the pivot rather than a fact about the applicant.

Run the same operation on Anneliese Vogt, a procurement director in Stuttgart with counsel, standing, and the option to walk away, whose carefully constructed conditional arrives at the other end of a machine-translated contract term as a commitment. The operation flags the same class of loss. She loses money and a negotiating position, and she can afford both, and the mechanism that took them is identical to the mechanism operating in Newark. The method is universal. The urgency is differential.

The Question That Stopped Waiting

#

Philosophy could once afford an open question. The relation between language and thought stayed unsettled for centuries at no operational cost, because nothing was standing on it. Departments could disagree. Positions could be held for a lifetime and passed on unresolved. Nobody’s asylum claim turned on it.

A question can stay open indefinitely as long as nothing is built on top of it, and something is now built on top of this one.

The infrastructure runs. Meaning crosses linguistic boundaries through a middle that nobody specified, nobody voted on, and nobody can inspect. The disciplines that would adjudicate what that middle is are downstream of a deployment they did not authorize and are not equipped to evaluate at the speed it is spreading.

Note what is not being claimed. The pivot is not wrong. It works, it works well, and it does something no designed system managed in forty years of trying. The category it operates in, sentence content, is a real category that captures something real. It is insufficient for what is being asked of it, and the insufficiency is structural rather than a defect, and no better version of the same thing closes the gap, because the gap is between what the representation is organized around and what a speaker is doing when they speak.

That is the stratum gap in its purest available form. The empirical layer is what the sentence says. The real is a person marking, in the only way their language provides, that they did not see the thing they are reporting.

The Notebook Again

#

Nasrin has nineteen years of margins. She has never been asked for them. The transcript is the record and the margin is not, and this was true before any of the infrastructure described here existed, which is worth saying plainly: the pivot did not create the problem. It industrialized a loss the institution had already decided not to count.

The hearings continue. The interpreters are being scheduled. The systems are being deployed into the sessions where no interpreter was available, and they are better than nothing, which is the comparison being made and it is the correct comparison. The middle is still unspecified. The question is still not being asked.

The notebook is full. Nobody has ever asked to see it.

This is the sixth essay in The Insufficient. It extends the series’ method from institutional categories to a representational layer, and it is the first essay in the series whose subject is a component rather than a practice. The four-hundred-year prologue to the universal-language dream belongs to Pax Interlingua, which spends it; this essay begins where the designed attempts failed. It should be read alongside The Missing Model, which names the account that was never built, and The Epistemic Framework, which specifies what an epistemic system owes its users. The indexical audit joins the seven operations of The Traditions as an eighth. That essay stated seven and treated them as complete, which was correct for what they were aimed at, categories applied to a life. This one is aimed at a representational layer instead, which is why it extends the set rather than revising the count. Like the seven, it detects and does not repair.

References

#

Designed Interlinguas and Their Failure

Hutchins, W. John, and Harold L. Somers. An Introduction to Machine Translation. Academic Press, 1992.

Vauquois, Bernard. “A Survey of Formal Grammars and Algorithms for Recognition and Transformation in Machine Translation.” Proceedings of the IFIP Congress, 1968, pp. 254-260.

Wilkins, John. An Essay Towards a Real Character, and a Philosophical Language. Royal Society, 1668.

The Emergent Middle

Artetxe, Mikel, and Holger Schwenk. “Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond.” Transactions of the Association for Computational Linguistics, vol. 7, 2019, pp. 597-610.

Conneau, Alexis, et al. “Unsupervised Cross-Lingual Representation Learning at Scale.” Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 2020, pp. 8440-8451.

Johnson, Melvin, et al. “Google’s Multilingual Neural Machine Translation System: Enabling Zero-Shot Translation.” Transactions of the Association for Computational Linguistics, vol. 5, 2017, pp. 339-351.

Meaning, Indeterminacy, and Grounding

Fodor, Jerry A. The Language of Thought. Harvard University Press, 1975.

Harnad, Stevan. “The Symbol Grounding Problem.” Physica D: Nonlinear Phenomena, vol. 42, no. 1-3, 1990, pp. 335-346.

Quine, W. V. O. Word and Object. MIT Press, 1960.

What an Utterance Does

Hymes, Dell. “On Communicative Competence.” Sociolinguistics, edited by J. B. Pride and Janet Holmes, Penguin, 1972, pp. 269-293.

Irvine, Judith T., and Susan Gal. “Language Ideology and Linguistic Differentiation.” Regimes of Language: Ideologies, Polities, and Identities, edited by Paul V. Kroskrity, School of American Research Press, 2000, pp. 35-83.

Silverstein, Michael. “Shifters, Linguistic Categories, and Cultural Description.” Meaning in Anthropology, edited by Keith H. Basso and Henry A. Selby, University of New Mexico Press, 1976, pp. 11-55.

The Disciplined Adjacent Literature

François, Alexandre. “Semantic Maps and the Typology of Colexification.” From Polysemy to Semantic Change, edited by Martine Vanhove, John Benjamins, 2008, pp. 163-215.

Haspelmath, Martin. “The Geometry of Grammatical Meaning: Semantic Maps and Cross-Linguistic Comparison.” The New Psychology of Language, vol. 2, edited by Michael Tomasello, Erlbaum, 2003, pp. 211-242.

Stratified Ontology

Bhaskar, Roy. A Realist Theory of Science. Leeds Books, 1975.

Series Anchors

The Approximate Mind, TAM-074 (The Interrogator): what can and cannot be asked of a system.

The Approximate Mind, TAM-075 (The Epistemic Framework): the specification for a system that reports its own limits.

The Approximate Mind, TAM-078 (The Missing Model): the account that was never built.

The Approximate Mind, TAM-102 (Pax Interlingua): the four-hundred-year dream, and the language the machines are building for each other.

The Approximate Mind, TAM-INS.02 (The Traditions): the seven operations the indexical audit joins.