Poetry's Resistance to Artificial Intelligence
The Machine Stops

By Will Stone

In his landmark collection Poems of Paul Celan, (Anvil Press, 1988) poet and translator Michael Hamburger includes a valuable insight into his modus operandi with the text On Translating Celan. He opens with the following truth; ‘the translation of poems includes two distinct functions and processes, which, for simplicity’s sake, I call reading and writing. By reading I mean everything to do with the taking up of the original text, from a merely intuitive grasp of its structural quiddity to a more conscious grappling with any semantic or referential difficulties it may present. By writing I mean the capacity to reconstitute the text in another language.’ He goes on to explain that almost all the perennial debates about literary translation are linked to this ‘delicate balance’ of the two disciplines, and how they have been shaped by the individualism of writers and translators down the ages as well as by the tides and currents of history. Poetry translation then is much more than just finding the right word for meaning. The machine faces a mountain and the slopes are of unstable scree.
 
Reading and writing as creative activities are then the two main arteries serving the translation of poetry, which serves to extend the original’s artistic potentiality. They are the widest bands detected in the strata of mankind’s achievement in history, literature and art over millennia. Even if a machine could somehow absorb all this human cerebral effort as ‘information’ and assessing it arrive at the most advantageous decision on a translated word, it cannot do so authentically or persuasively, due to the infinite vagaries, twists and turns, circuitous routes and dead ends of the wholly human paths that converge on that individual brain of the human translator hunched over his writing desk, in whichever era. Such a path invariably includes those spanners in the works named chance, absurdity, and even more importantly heroic failure, a valuable human quality in terms of the development of culture but one which technology treats in only binary fashion, as something to be overcome not celebrated.
 
But surely any enquiry into the prospects for Neural Machine Translation (NMT) to tackle the art of poetry translation, must pose the key question as to whether it is to be solely a tool of the human literary translator, as currently, or to be his or her replacement, a future doomsday scenario. Well-meaning and able academics and scholars in the field of translation studies, tend to skate spiritedly around this apocalyptic elephant in the room, either losing themselves in the technological wizardry of NMT or seizing on the positives in the ways that NMT can play a support role, a secretly ambitious cyborg apprentice who still needs the master to correct his raw proofs. As yet he doesn’t quite have the fine finishing qualities of the master’s brush, but he is learning…
 
Literary translation is an act of creation, the translator being a bona fide writer, not just a plodding lime kiln donkey carrying words back and forth between languages. But because translation is also an act of ‘rewriting’ and craftsmanship after the event and involves a physical interchange of language, it is much more vulnerable to the machine. It does not possess that sense of the sacrosanct, the regal air that blows around the original work. Furthermore, literary translation, although a heavily trod crossing point, is a temporary structure and can be reshaped. It is vulnerable in a way the original is not. A literary translation helps the original survive and prosper, it leads, but the guide must be familiar with the terrain between languages. He or she must have passed that way before and the led must have confidence in the guide. The machine may be able to scan the landscape and assess the topography but it hasn’t actually walked the route, it hasn’t felt the landscape with all the senses, it hasn’t held the sound of the word, reflected on it, it cannot judge tone with an actual ear, nor reconstruct the resonance that human vocal cords will provide during a reading. Some may claim this does not matter, but poetry stands and falls on its music, in a way most modern prose doesn’t.
 
Perhaps only poetry’s shadow-play between languages can create an impasse for NMT, prose is an already conquered territory. Like those migratory birds who remember their routes, NMT recognises word and sentence patterns and logs them. But poetry is light-footed, elusive, irregular, more is compressed into less space, the mighty visionary image rises out of limited but precisely crafted language. Language in poetry has a lot more to do and little space in which to do it. The risks are greater. I have had experience of translating both prose and poetry. Prose is generally a sequence of sentences of varying length with consistent punctuation that (unless you are Thomas Bernhard) become paragraphs, then chapters at convenient pausing points. A sentence of prose has a start and a completion; between those bookends description, scene setting, action takes place. The machine learns. But with poetry there are no sentences as such, intent is obscure, content is ambiguous and placements strange. Poems are like those plants which do not fill the available space, but rather keep to their own, they are anchorites, sequestered from the vast prairies of prose texts where thousands of hooves thunder. With the poem things slow down, reflection is paramount, the mind must go inwards, away from the language in a sense and towards the image, the pictorial. Poetry with its shorter lines, which might drop off at irregular points, the eccentric structures and non-conformity inherent to the form, the preponderance of images, visions, rampant metaphor, the switching between symbolism and reality within a poem, within even a line of poetry, must surely trip up technology. The ability to read the whole, not just the line seen through a narrow visor, affects the human translator’s textual decision. The machine may reap industriously in the field of verse translation, making its calculations in a nano-second, yet it cannot ever complete the harvest. Even the poet and subsequent reader of their work may struggle to access its first layer of truth, since the poem is constantly transforming itself across different cultures, eagerly welcoming interpretation.
 
Walter Benjamin states in his The Task of the Translator, an invaluable text dating from 1923, ‘the task of the translator consists in finding that intended effect upon the language into which he is translating which produces in it an echo of the original.’  An echo, not a mirror image. Benjamin goes on to champion not works of literature themselves but their translation as the ultimate language of truth, he suggests a third language, hovering beyond the other two, the translation restocks or ‘energizes’ the original, so each language is ‘supplemented and reconciled in their mode of signification’. For Benjamin translation is the great unifier, the extender, harmonizer, repository and shepherd of individual languages. ‘Unlike a work of literature, translation does not find itself in the centre of the language forest, but on the outside facing the wooded ridge; it calls into it without entering, aiming at that single spot where the echo is able to give, in its own language, the reverberation of the work in the alien one.’ But the Neural Machine interminably wanders the language forest, lost in a labyrinth of possibilities, unable to draw back from the goal of finding the most accurate meaning, unaware of the commanding aura Benjamin attempts to articulate.   
 
In more practical terms, human versus machine, the following poem may serve as an admittedly insufficient example. The translation is by myself and taken from the series Poems to Night by Rainer Maria Rilke (Pushkin Press, 2020). Below it is the available google translate version. Although the machine appears to impress as regards the meaning of the lines, it is unable to harness the stylistic skills which a translator of poetry customarily employs, such as alliteration, assonance, near rhymes, end rhymes and so on. These are the means by which a translator can ‘compensate’ for what is lost in the inevitable challenge of bringing across the rhythm, the sound or music, call it what you will, of the original as defined by its semantic structure, its formal rhyme scheme, i.e. elements inherent to the original language within the poem’s structure, influenced by the style choice of the author, the era in which they lived. But these re-engineering methods are best used not as specific ‘tools’, but as an unconscious element during the writing of the ‘new’ poem, which is why being a poet is highly advantageous though not a prerequisite. They should be gently coaxed but not cajoled.
 
In the very first line the German word ‘verschwendeter’ poses a problem, for it has a variety of associated meanings including ‘wasted’, the machine’s choice, and ‘squandered’, the human’s choice. Although ‘wasted stars’ seems appealing at first glance, I opted for ‘squandered’, since I felt that the poet was implying that the human onlooker had not appreciated them, (we learn this further on in the poem), so yes indeed ‘wasted’ them, but ‘wasted’ has obviously another connotation, that they are dying, denuded of strength, drawing us to the false friend of the falling, dying star, but I do not see that here. Supporting  the case for ‘squandered’ is its alliterative liaison with both ‘skies’ ,‘stars’ and also ‘splendour’ in the next line. Although Rilke’s poem is not profuse with rhymes internal, end or otherwise, we have ‘Kümmernis’ and ‘Antlitz’, ‘weine’ and ‘weinenden’, ‘greifend’ and ‘hinreißend’. Apart from the mileage given by ‘space’ and ‘face’, machine and human were hard pressed to echo these sounds with the English available and within the remit of meaning. At what point should the translator risk breaching the meaning contour of Rilke’s image to produce a more tonally ‘respectful’ echo?