Geronimo! Neural machine translation, post-editing, and the post-human

by Rafaël Newman

Notwithstanding the spread of English as a global lingua franca, translation continues to be a vital component of international relations, whether political, commercial, or cultural. In certain cases, translation is also necessary nationally, for instance in countries comprising more than one significant linguistic group. This is so in Switzerland, which voted by an overwhelming majority in 1938 to add a fourth national tongue to thwart the irredentist aspirations of its Italian neighbor, and which in certain contexts is obliged to use a Latin version of its own name (Confoederatio Helvetica) to avoid favoring one language group over another.

With its four languages, three of them national and official – German, French, and Italian – and the fourth, Romansh, “merely” national, Switzerland is indeed obliged to do a great deal of translation, especially at the level of its federal ministries and law courts. Its commercial enterprises, too, typically depend on communication in at least one other language region than their own immediate location; and naturally, many Swiss businesses have a linguistically diverse national presence in any case, and thus require a polyglot corporate identity.

Culturally, although Switzerland’s linguistic regions tend to look to the “motherland” of their respective language in matters of tradition, the country’s creative class is by necessity international in its outlook, given the limited size of its domestic market; while its chief cultural funding agency, Pro Helvetia (bearer of a similarly non-partisan Latin appellation), spends a great deal of its resources on representation and “localization” abroad. And finally, since Switzerland’s economy is strongly geared to export, and because its lack of natural resources means that it has come to specialize in services and end manufacturing, those sectors, particularly the financial and pharmaceutical branches, are positively ravenous consumers of translation services, especially into the global tongues: Chinese, Spanish, and of course, above all, English.

No surprise, then, that those same sectors are presently lured by the cost-cutting Siren song of translation “solutions” based on artificial intelligence, whether for their in-house language services or from the agencies to which they outsource their translation orders.

The professional translation business, long adapted to the use of computer-assisted translation (CAT) tools, is undergoing a revolution at the hands of machine translation programs using so-called artificial neural networks, which mimic the operation of their biological namesakes and are able to “learn” based on data input during extensive “training” sessions. This new technology promises up to 30% increases in productivity (with anecdotal evidence suggesting that greater gains are possible), which means significant financial savings given the volume of text processing by larger firms, and the relative cost of the relevant human resources, AKA translators. And, although Switzerland, traditionally slow to evolve, is well down on the list of first-adopter countries, its major service providers are currently competing with one another to develop their own bespoke neural networks, for the proprietary treatment of material often containing sensitive client data, and thus find themselves, as it were, fighting for a seat at the very front of the last car on the train. And yelling “Geronimo!” as they sprint down the platform, coat-tails flying.

What gladdens the hearts of chief financial officers, however, is not being met with universal enthusiasm among professional translators of “pragmatic” or non-literary texts; indeed, in some cases, the introduction of AI is handled with positively McKinseyan circumspection by “change managers,” whose brief is to prevent panic. And no wonder, with translators’ jobs increasingly becoming a matter of what is known euphemistically (and pleonastically) as “post-editing”, but is more trenchantly described as “cleaning up after the robots we have schooled to replace us”.

The process is as follows: a text to be translated is delivered in a standard format (Word, PowerPoint, Excel, etc.) and is then “opened” in a CAT tool (such as Trados Studio) and thereby transformed into simple text with formatting tags. The text is further analysed into “source segments”, sentences or fragments terminating in variously predefined marks (periods, colons, paragraph breaks), arrayed vertically down the left-hand side of the program window, with blanks in a corresponding column to the right, ready to receive “target segments”. These last are then completed in various fashions. The translator, operating with good old-fashioned linguistic knowledge and cultural literacy, can turn a source utterance into a target counterpart. The CAT tool itself may consult the “translation memory” linked to it, a database comprising all of that translator’s (or their team’s) previous translations of texts, to choose matches of a predetermined statistical accuracy (typically 80% and above), for subsequent review and revision, as necessary, by the translator. Or the artificial neural network, also connected to the tool, operating according to a process known as “neural machine translation”, or NMT, may determine the best possible translation of a given segment; it does this based on prior assignment of numerical coordinates to linguistic elements, and lightning calculation of their likelihood as an equivalent according to “learned” relationships of semantic proximity.

This is the standard scenario for in-house translators, who receive a salary and, while obliged to adhere to productivity targets and daily deadlines, may choose to accept or reject the proposals made statistically, by the database, or neurally, by the network. External contractors, for their part, who are paid by the line, receive their “project packages” – converted CAT files with attached memories – in “pre-translated” state, which means that the segments have all been filled in automatically in advance, whether with statistical or neural proposals, in the course of a process triggered by an account manager; and their work – to the extent that they still receive any, since part of the aim of digitalization is eventually to do away with their putatively more expensive services – will thus consist essentially in correcting the algorithm’s stylistic infelicities and outright mistakes.

It is these last – the “mistakes” made by NMT – that give post-editing, which is otherwise drudgery, a certain philosophical interest: because the errors generated by NMT are qualitatively different from those made by the “repair” function built into the statistical database linked to the standard CAT tool, which routinely attempts to make up for a deficiency of 5 percentage points in a potential source-target match by helpfully suggesting absurd inclusions (such as “hat trick” or “DAS Automotive Insurance” for English translations of German sentences containing novel uses of the third person singular of the verb “to have”, or a subordinate clause involving a neuter noun). Mistakes made by NMT are often subtle, involving omissions or reduplications reminiscent of the errors made by medieval scribes, for which scholars have evolved a specialized vocabulary: these moments of inattention are known as “haplographies” and “diplographies”. NMT errors may also display ostensibly cunning inventiveness, as when an English product name in a German context is translated, in an English target segment, into a third language, learned apparently by cross-pollination from a parallel linguistic pair. Or they may go off the rails altogether, resisting interpretation by producing Dadaist sequences of nonsense syllables when confronted by an ambiguous or un-analysable construction – “Zinseszins”, the otherwise common-or-garden German term for compound interest, unaccountably mutating to resemble the name of a disgraced French footballer: Zinazinadinadina

“A man of genius makes no mistakes,” says Stephen Dedalus, Joyce’s alter ego. “His errors are volitional and are the portals of discovery.” The mistakes made by NMT, requiring post-editing by professional human translators, are not especially genius, do not usually provide access to discovery, and cannot by any standard criteria be termed volitional; but they do constitute a quandary for their correctors, not only as professional translators, but more urgently as human beings. For language, after all, is what sets us most recognizably apart from other animals. It has been variously identified as hardwired (Chomsky) or deified as that which speaks us (Heidegger). Freud, of course, and more explicitly his apostle Lacan, insisted on the linguistic structure of the unconscious, and “read” its communications precisely at such inconvenient junctures as mistakes, both of spoken and of written language. (Although the Marxist philologist Sebastiano Timpanaro argues for the purely mechanical nature of the scribal error as a retort to Freud’s theory of parapraxes as evidence of unconscious desire.) And as for translation, George Steiner has proposed it as the activity subtending all acts of human interpretation, whether inter-lingual or not. Furthermore, the ability to innovate within a heteronomous semiotic tradition is the essence of poetry. It is what humans do in excess of what is strictly necessary for survival, purely in pursuit of “self-enjoyment” (although Hegel maintains that birdsong also represents such an activity); it is our means of marking the boundary between ourselves and other species; and it is the mechanism of symbolic individuation. “Children become people when they wriggle out of your arms and say ‘no’,” writes Olga Tokarczuk, whose novel of non-human resistance, Drive Your Plow Across the Bones of the Dead, proves the rule of humanity’s monopoly on protest in the course of suggesting the exception.

Here, however, is a machine – it’s in the name, neural MACHINE translation – that can translate, err creatively, and resist. In an eerie reprise of the factory mechanization that revolutionized modern industry in the early 20th century, it is no longer the human workers who are becoming machines; it is the machines that are taking on human traits. What will distinguish us in future from the robots we are ceding our functions to? What will remain the shibboleth of humanness? Is it the mundane ability to recognize discrete elements of a fragmented sign system in pursuit of a commercial transaction, as in the test proposed by online payment portals (“Confirm you are human: click on all the squares showing traffic lights”)? If it is the ability to dissemble, and thus to alter reality by inventing narrative, then why has Facebook found it necessary to prohibit the dissemination of content that is “the product of artificial intelligence or machine learning that merges, replaces, or superimposes content into a video in a manner that makes it appear authentic”? Perhaps it is our ability to experience, and elicit, sensations of love and tenderness: and yet robots are being devised to provide comfort across the entire spectrum of affections, from agapé to eros. The novelist Ian McEwan’s last-but-one, Machines Like Me, involves the dystopian fantasy of an AI-driven creature named Adam developing feelings for the narrator’s girlfriend; when the two consummate their relationship within earshot of the protagonist, the last finds solace in the thought that Adam is nothing more than an outsized vibrator – which does not, however, stop him taking all-too-human revenge when Adam begins to exhibit inklings of an all-too-human perversity.

Perhaps the last bastion of humanity is in fact simply such perversity, which the gender theorist Maggie Nelson, in an Aristophanic move, nominates as the symbolon of a romantic relationship: find the person who shares your particular perversity, and Bob’s your uncle/aunt. (Nelson also writes compellingly about the sheer absurdity of the fact that pregnancy, a resolutely “queer” experience of reduplicated embodiment and self-estrangement, is deployed as a heteronormative paragon.) Our tendency to act against the dictates and norms of reason, of tradition, indeed of self-interest, in the pursuit of an idiosyncratic aim gives rise to art, to culture, to religion, as well as to war and depredation.

By definition, perversity manifests endless variety, on both the individual and the societal level; as a society, we are currently indulging our perverse taste for progress by any means necessary by automating and digitizing away entire categories of service labor, from cashiers to caregivers. We are driving the ranks of desk workers, such as text processors, editors, and translators, into the precariat, in the name of abolishing allegedly “dehumanizing” forms of work. According to democratic socialists like Martin Hägglund, however, what such a move also requires is a radical re-evaluation of leisure time, granting it the centrality heretofore enjoyed by labor and making it possible for the ranks of the newly under-employed to make meaningful use of their non-working time. Instead, as we plunge headlong into a future of exponential population growth, global climate change, an increasingly robotized and downsized labor force, and dwindling resources, trained by change managers, austerity politicians, chief financial officers, and other McKinsey graduates to believe in the inevitability of our own redundancy, we ourselves enthusiastically take up their cry: “Geronimo!”

Which last is itself a veritable palimpsest of human perversity. First used by US army paratroopers in the 1940s aping a Hollywood historical fiction, “Geronimo!” has since become a standard battle cry among American soldiers. The expostulation echoes the westernized sobriquet of one Goyaałé, an Apache resistance fighter who is said to have himself cried “Geronimo!” when fleeing US troops; it belongs to a repertory of native American customs adopted by European colonizers in a cannibalistic gesture of domination, and thus recapitulates in nucleo the infamous history of violence that subtends all manifestations of human progress.

The name “Geronimo”, of course, has another, older history. It is the Romance version of the Greco-Latin Hieronymus, “holy name”, handed down to English by way of French as “Jerome” and best known for its 4th-century bearer, Eusebius Sophronius Hieronymus, or Saint Jerome. A priest and theologian, Jerome is celebrated for his Latin rendering of the Bible, a version now known as the Vulgate. Caravaggio painted the iconic image of the saint in the early modern period: seated at a desk covered in books, Jerome sports a halo, the sign of his otherworldly destiny, while his work is overseen by a skull perched atop an open volume, a memento mori reminiscent of this world and the presence of death among the living, as well as of the gravedigger scene in Hamlet, a contemporary work. As he innovates within an ancient tradition, in other words, and transports holy texts from a foreign past into a domestic present, Jerome is at once innocently illuminated by a future not of his making, while remaining thoroughly aware of a past strewn with the bones of his mortal forebears.

Naturally, he is the patron saint of translators.