Poetry // [Machine] Translation

Everything you are about to read is based on nothing.

[ Updated: 8/Aug/2018 9:00AM ]

If you are still reading this you are conducting an illogical act. You were just handed a piece of information suggesting that the following article is unfounded, may carry unsubstantiated information or fallacies, and is generally unsupported by data. Yet, you are still here. Which is exactly what you do when you read a poem.

Poems, Poems

The oldest surviving poem is thought to be the Tale of the Shipwrecked Sailor, written in Hieratic and ascribed a date around 2500 B.C.E. Other sources differ, ascribing the earliest written poetry to the Epic of Gilgamesh written in cuneiform. Not taking sides, it is clear that poetry, especially oral one predating any writing system, is probably as old as language itself.

But what is a poem? It would be unwise to discuss poetic translation without first trying to grasp the essence of poetry itself.

Looking for a formal definition one finds that a poem is: “A piece of writing that partakes of the nature of both speech and song that is nearly always rhythmical, usually metaphorical, and often exhibits such formal elements as meter, rhyme, and stanzaic structure.”

Sounds simple enough. The problem is that the former definition places nursery rhymes alongside Wordsworth  on a single metaphorical bookshelf. It is too wide to be of any practical use.

Maybe poems defy definition altogether, however denying poems can be defined is a futile experiment in solipsism that suffers from the same drawbacks as its prior definition, and is too, practically unusable.

“Poetry is that which is lost in translation”
Robert Frost

It is no accident that poems evade formal definitions. They command a multitude of features, some lingual, some harmonic and temporal, an undertone of verse and rhythm, visual or morphological cues and more.

Some poems, simplistic and wry as they may be, trigger emotion in ways so fleeting they do not easily fall into any of the above definitions. Trying to map all the different forms of poetry is a multi-dimensional problem, where dimensions heavily interact until they are practically indiscernible.

Poems are a holistic experience. Reading one, you are drawn into a mental image of your own creation, a trip down imagination lane. Each reader reads it differently, in his own tone and tempo, creating his own images and emotional responses. The same poem, when read by different people of different cultures may trigger completely different meanings and emotions.

Take a minute to read All The Way by Charles Bukowski. Mind your internal voice, tone and intonation, where you pause for a brief moment, where you carry through. Register you visual reaction to the text layout, and your emotional responses and how they fluctuate.


Now listen to Bukowski reciting his own poem.

Was it any different? Completely different? Did you get any insights by listening to his version as opposed to how you read it? Keep those insights and observations in mind, they would soon become useful.

Can Poetry Be Translated

I don’t mean can it be machine translated. I am asking a prior question, a more theoretical one. Do you think that poems can be translated from one language to the other while keeping the exact original overall meaning?

The obvious answer is: yes. After all, poems are and have been translated. Hundreds of thousands of poetry books serve as an overwhelming evidence.

Do they?

Be careful. The fact there are textual representations of the same poem in different languages does not support the above claim. Poems may be translated and printed in languages other than the original, but rarely does the translator succeeds in keeping the entire set of features of the original poem. That is not due the translator lack of proficiency, rather because it is not theoretically possible.

A single poem read by two different people is read differently. In case those two people originate from different cultural backgrounds and mother tongues it is even more so. Even a single person, reading the same poem on different days of the week, or periods in life, may experience different insights and emotional responses.

The Process

At this point I hope you agree that an exact translation of poems is a theoretical impossibility. However, poems do get translated.

What is the process of poetic translation then? How does a human translator, proficient in both source and target languages, translates the poem from source to target? The answer is: she does not.

Not because she does not produce a target translation, she does, rather due to the fact that the process she employs has nothing to do with what is commonly perceived as translation.

When a human translator takes-in a poem in the source language, she first grokes the entire set of multidimensional features conveyed by the poem. The translator must deeply understand the nuances, cultural references, word fusions if any, textual structure, rhyme and tempo, and many other features unique to that specific poet and poem. There is also no escape experiencing the emotional effect conveyed by the poem itself. Only once the translator internalized the full set of interplaying features, does she have a comprehensive mental model of the poem.

We do not yet know how this mental model is arranged in the brain, but we can safely assume it exists and that it is distributed over many areas of the brain. If you are still not convinced try reading a poem. The sensation is unmistakable, the feeling of multidimensional understanding of the music of the words combined with their meaning and the concurrent excitation of emotion cannot be denied.


Then a miracle happens. Kind of.

The translator becomes a poet.

He creates a whole new poem in the target language while keeping all the source poem’s constraints. The target poem must convey all (mental) model features and their relationships, as represented in the translator brain neurology.

The process is similar (although much more complex) to the mathematical concept of constrained optimization. In mathematics we try to optimize an objective function with respect to some variables in the presence of constraints on those variables. In poetic translation we try to generate a set of strings (words), given a set of constraints those words must obey (e.g. rhyme, meaning, visual layout, etc.).

At this point you should correctly assert that all translations of some poem done by different translators are in fact newly created different works of art, collectively and individually unique. Starting from a single source-poem we now have N new ones, each an artistic utterance of its own.

Machine Translation

The problem of automated poetry translation has been around for some time. Researches from major companies and academia have been trying different machine learning and AI techniques in order to provide at least a decent translation of poetry.

It proved to be a very hard problem.

Most translation software works by searching through a wealth of possible translations, then evaluating what’s most accurate.

Whereas, “if you translate poetry,” Genzel says, “you have to preserve what you want the reader to feel.”

Although some poets insist even human-to-human poetry translation is incapable of matching that “feeling,” Genzel’s research does take some big steps toward preserving a poem’s length, meter and rhyme.

Translating a haiku? Genzel can preprogram his computer to generate online lines of five, seven and five syllables.

A Shakespeare sonnet in iambic pentameter? The computer can read a pronunciation dictionary, Genzel says, “like you would use to learn another language.” Once it knows where the stress falls in a given word, it can correctly place that word in a metered sentence.

“The hardest thing to do is rhyme,” Genzel says, “because it connects to different places in a sentence,” and because two words that rhyme in English may not rhyme in another language.

In that case, the best Google can do is to cycle through a long list of optional matches to find a rhyme that’s right.

2011, “AI Complete” project, Dmitriy Genzel, a research scientist at Google

Genzel has done some progress, especially when the poem obeys some well-known stanza. Less, as generic poetic translation goes. And even less so when rhyme is concerned (see above).

A more recent study (2013)  begins with the following statements:

“Translation is a complex, multifaceted challenge. It aims to produce an equivalent text in another language, but this equivalence is difficult to define. The translated text should communicate the same meaning, the same picture on a mental stage. It should match the style and structure of the original. Translation demands nuance and detail, a subtle, artistic pursuit.

Poetry is particularly difficult to translate, as meaning and form are intricately interwoven. Rhythm and sound play a key role in many poems, and an effective translation will capture these effects. It seems impossible for machines to capture the artistry of language in even mundane documents. Yet machine translation systems continue to improve at translating a variety texts. However, these systems focus on preserving meaning, not capturing the rhythm and sound qualities important in poetry.” [2013, Poetic Machine Translation, Girardeue and Rajpurkar]

Girardeue correctly identifies the multi-dimensionality of poetic expression, suggesting a way of mechanically dealing with such complexity:

“We propose to add a poetic model to a statistical machine translation system that evaluates the poetic or lyrical quality of a translation. This will cause the system to find more poetic translations by heavily rewarding those that satisfy a given poetic form. There are a variety of potential components in such a model, with the most relevant (in English), stress pattern, rhyme, and syllable count.”

Sadly, even such advanced models are limited to a small number of poetic and stanza structures, alongside a small number of features. Notably, it is Girardeue himself who admits: “It seems impossible for machines to capture the artistry of language in even mundane documents”.

Some Very Bad Math

When computer scientists cannot provide a robust algorithm solving some problem, they usually settle for the next best thing. Proving it cannot be done.

Let us try and shatter any theoretical possibility of ever achieving full and exact poetic translation. To do so we need to call on a dead mathematical giant; Kurt Gödel.

In 1931 Gödel shook the mathematical world publishing his two incompleteness theorems when he was only 25 years old. The first incompleteness theorem states that no consistent system of axioms whose theorems can be listed by an effective procedure (i.e., an algorithm) is capable of proving all truths about the arithmetic of the natural numbers.

Or as Douglas Hofstadter described it in his extensive book Gödel, Escher, Bach:


We can then use Gödel’s theorem to show that there could never be a computer program (algorithm) that could translate all poems.

To do so we replace axioms by syntax rules, theorems by poems and the notion of proof with that of generation (i.e. a process that builds a new poem out of an exiting one) to immediately get:

No consistent system of syntax rules whose poems can be listed by an effective procedure (i.e., an algorithm) is capable of generating all target poems.

The pseudo mathematical construction stated above can be paraphrased as:

There would always be a poem that could not be translated by an algorithm.

This of course is extremely bad math.

One cannot simply abuse a mathematical theorem whose proof is (beautifully) tailored to number theory, freely transforming it to apply to strings (words) and lists of strings (poems). However, it points at one possible direction should one like to explore applying the notion of incompleteness to the poetic machine translation problem [Let me urge you to provide such a proof – DS].

The Death of the Translator

And maybe it’s the other way round. Professor Tong King Lee, of the Hong Kong Polytechnic University published a paper titled: “The death of the translator in machine translation, A bilingual poetry project” (2011). 

Lee’s paper explores the notion of the death of the Translator, inspired by Barthes’ formulation of the death of the Author. It argues that the death of the Translator is caused by a loss of human agency in translation and is therefore most clearly exemplified in machine translation.

The research was based on an avant-garde bilingual poetry project by a Taiwanese poet [I have requested Prof. Lee for his paper and would extend on this project here as soon as he provides it – DS], the paper demonstrates that machine translation can produce unexpected new meanings through unpredictable routes of semantic and syntactic divergences from the source text. Or in layman terms, machines generating poetry out of existing poetic texts are being original in ways we could not understand.

As Barthes claimed, “it is language which speaks, not the author”, meaning the machine interpretation of the poetic language as exemplified in what we (wrongly) call a bad translation, is nothing but. It is in fact an original interpretation, much like that of the human translator generating a new poem based on her mental model of the source.

In a sense the machine and the person, performing the act we refer to as “translation”, are following the same process. The only difference is in their respective interpretations. As the machine’s internal language and representation differs from that of the human, it simply views the poem in terms the latter cannot begin to understand, thus generating a coherent poetic translation in a lingual sphere we can never truly be a part of.

In the summer of 2017 the tabloids cried: “Facebook AI Creates Its Own Language In Creepy Preview Of Our Potential Future” [Forbes, 2017]. What actually transpired was that Facebook’s engineers have been trying to develop AI that would make digital communication more efficient. As a result of that the chatbots they were using for the task developed their own representation of language, which as one must admit was mostly gibberish. As opposed to the media outcry, those chatbots are not yet taking over the world, communicating in a language we cannot understand, but it makes one think. Do we really own language, or is our concept of language merely one interpretation amongst many.

It’s The Brain, Stupid

A research done in Exeter university in 2013, used fMRI to record brain activity of subjects reading poetry. Results show multiple areas of the brain lighting up simultaneously, including textual areas, areas that ordinarily respond to music alongside emotion related ones. This research supports our claim that poetry is a holistic-integrative experience switching-on multiple interconnected modalities of the brain.

To do poetic machine translation right we must first create a much deeper and systemic theory of the brain. Its integrative reaction to poetry and models accurately describing emotional response. Only then can we capture the elusive sense of poetry. However, apart from a recent paper titled R17 [2017, Short Version of ‘The Brain, Explained’:
A Response Process Theory of Brain Function, Rappoport] current research is surprisingly lacking.

Once we acquire that level of understanding it is not unfathomable that machines may be able to not only translate poetry as well as humans, but also generate original, emotionally-moving verses of their own.

With mankind watching, the World President said, “Fifty years ago, you were declared The Sesquicentennial Robot, Andrew.” After a pause, and in a more solemn tone, he continued, “Today we declare you The Bicentennial Man, Mr. Martin.”

Isaac Asimov,  Bicentennial Man