Machine Translation of Poetry:
Current Practices and Constraints

“A poem is a small (or large) machine made of words,”
─ William Carlos Williams

Translating poetry is often a tricky task for human translators, and it is believed that machines are not capable of doing it. Machine translation of poetry is considered to be the most difficult task not only in computational linguistics but also in artificial intelligence in general. It is generally believed that involving computers to translate verses is a very rotten way of writing poems. But what happens if you create a poem with a computer as a partner? Such collaboration is a subject in Charles Hartman’s book Virtual Muse. The author argues that artificial intelligence will eventually equal human brain in writing and translating poems (Hartman, 1996). At present state of machine translation, practitioners often have to choose between getting either the form or the meaning right; however, preserving both is controversial and constraining. Vladimir Nabokov in his translation of Eugene Onegin disparages the translators who are trying to preserve the form, arguing that the form should be sacrificed for the sake of meaning. Others believe that the form should be preserved in order to maintain the sound and the meaning of a poem. Computers seem to be very well-suited for such tasks because in poetry there is no conformed standard; therefore, one should not be afraid that a machine-produced verse will not meet certain poetic dogmas and canons. Also, when translating a poem manually, you have to edit it multiple times; however, there is no principle difference between editing your own “verse” and the one that machine produces for you. This paper hypothesizes what barriers to the machine translation of poetry are most likely to be overcome by computational advances as well as how to evaluate computer poetry based on translation algorithm type and cognitive processes in translator’s mind. The article also comments on what gets lost in machine translation of poetry, what computers fail at and how human understanding, intuition, cultural references, sensitivity, and other constraints might impact computer-generated poems.
Poetry is often regarded as the highest expression of language and poses many challenges to translators. As someone said, “poetry” is what is lost in translation but the key is to allow the translated verse be a poem in its own right.” Since capturing rhyme and meter is rarely possible even in manual translation, one should allow certain liberties with the lines in machine translation – for the sake of the poem translated, for example, into English. Provided that the poem can exist on its own, the liberties can be introduced at word, grammar, and meter levels. Capturing the poem’s rhythm makes a lot of sense, especially when it has something to do with emphasis, i.e., having the stress on the key words. As far as machine translation goes, constructing grammatical and sensible utterances is not always possible, and, therefore, the above-mentioned liberties can be justified. Google Translate, for example, generates texts by searching through a huge database of possible translations, guided by a model of accuracy. In general, translation accuracy often has to be sacrificed for the sake of poetic form. A Google researcher Dmitriy Genzel argues that “succeeding in machine translation of poetry is possible and that it is not such an endeavor as people might think” (Genzel, 2010). Google is now designing the software that would be capable of translating a poem and retaining its rhyming scheme and meter or transforming the poem into another form – from limerick to haiku for example. In order to do this, Google is changing its algorithms (search procedures) to sacrifice some accuracy without, however, affecting the form. Since computers can analyze very large text repositories of human works and maintain these in memory, rhyming – being the hardest part – is possible, and the machine simply cycles through a long list of possible matches to find the most probable. In other words, the machine “digs out” rhyming matches from the existing online poetry corpora. The drawback of this innovation, as it has been mentioned above, is that the accuracy of meaning is often lost in translation. Although the machine picks the most statistically probable rhymes, they are not often accurate, and the rhyming words in the translated poem are rarely direct equivalents of the words in the source text. Moreover, the words in one language do not necessarily rhyme with the words in another language. What is pertinent to understand in preserving either form or meaning in the machine-translated poem? It is the pragmatic approach being taken: if we cannot render the message of a poem the way the original (author) requires, then we apply a pragmatic approach, i.e., focus on the reader’s perception, and, if necessary, preserve only rhyme and meter, allowing inaccurate lexicon as necessary. Any inaccuracy, however, can be fully or partially avoided provided that the translator edits the draft version of a computer-translated poem and duly “polishes” it. More pros of machine translation of poetry are further hypothesized in the next paragraph.
Can machine-translated verses be labeled as good quality poems? Many translation programs are of a fairly common kind, they are filters: you take some preexisting text and you pass it through the program that transforms it in some way and produces a translated text that you might find either interesting or gibberish. The vocabulary and syntax of such a text are often plausible, correct but randomized. The idea of randomness is crucial here. It can be helpful to a poet or translator to have some kind of language input that is startling, unexpected, something you could not have predicted. That is a very fascinating part of the process because the computer is actually giving you a translated material and then you are shaping it through editing and proofreading. That is the key feature of this relation, this collaborative experiment. Although it sounds like a weird way of writing poems, it does not cause any harm to try, mainly because it does not feel that much different from the rough poem drafts produced manually. Very often, when you sit down to translate a poem, you do not know where you are going, and, naturally, a lot of what you produce is silly, useless, and not productive. Then, after you have eventually produced some translation, you have to sit down and become an editor, and becoming an editor of your own translated verse is not at all different from becoming an editor of the material that the computer gives you. Poets often experiment with their vocabulary, then why cannot computers? What the computer, however, fails at drastically is reading, not writing, i.e., the machine cannot edit; it can only generate and produce (translate) things. The computer does not yet know how to look at what it has put out and decide whether it makes sense, decide whether it is interesting, decide whether the task of making meaning out of it, which any reader has to do, is worth doing or not.
On the one hand, the computer-produced “verse” does not make sense because the poem is supposed to come out of human brain. There is a mechanical underpinning to what the brain does, and the most fundamental difference between the computer and the brain is that, comparatively speaking, the machine is really unintelligent and the brain is really smart so far. On the other hand, one can make an experiment and see the difference: take a page generated by a computer program and give it to a group of students without telling them that the material was translated by the machine. As an example, Hartman suggests an extract from a book of poetry The Policeman’s Beard is Half - Constructed. The poem was generated by a program called Racter in 1984:
Bill sings to Sarah. Sarah sings to Bill. Perhaps they will do other
dangerous things together. They may eat lamb or stroke each
other. They may chant of their difficulties and their happiness.
They have love but they also have typewriters. That is interesting (2).
The above verse sounds pretty thought-provoking and marvelously funny. The students would have to do what every reader should do - find meaning. If they are good at it, they can make a lot of meanings out of the text that, actually, has no author. Then, if you tell them this was produced by a computer, their ranges of reaction can be quite surprising, from excitement to even madness; therefore, the computer can be seen not only as a tool but as an artificial intelligence. If artificial intelligence becomes equal with human brain while translating poetry, then we might not see any differences and to us it would be like talking to another person. A Russian computer programmer Sergey Teterin launched an interesting project Cyber-Pushkin 1.0 beta in 2003, a so-called prototype of artificial intelligence. He designed the software that works in MS DOS and generates poems using the memory compiled of thousands of Russian poems. Filled with the best works of Sergey Esenin and Aleksandr Pushkin, Cyber-Pushkin produces very likeable poems where rhyme is preserved. They sound gibberish; however, after editing, such poems can be labeled as pretty sensible.
Unfortunately, Google’s experiments seem to terrify many professional poets and literary translators. Many of them hold onto the notion that poems are not statistical problems to be solved by machines. Critics claim that in order to produce high quality translations of poetry, machines must possess the qualities of a human translator. Obviously, the machine’s English must be fluent so that to produce a clear output. There are many implications here since language should be seen as an active form of engagement with the world. Every word has echoes: pick a thesaurus and you can appreciate a richness clustering around every word or collocation; therefore, the machine must not only have access to dictionaries but also be capable of understanding equivalences, which is important for an accurate translation of poetry. Machine must be also competent in constant testing and updating words, forming patterns of oral, written, and vocal language. Also, the machine must understand not only meaning but intention, i.e., what underlies each poem, its author’s conscious and sub-conscious ideas. Provided that the machine learns the above skills, one can say it is conscious and ready to tackle poetry. If it becomes equal with human translator, then the evaluation of machine’s output should be the same. Judging the machine’s performance, one should consider the following: does the poem live and have integrity, does it add to knowledge of the original, and does it metamorphose that original? If it fails to handle cultural references, lacks sensitivity, or has no emotional impact, then the translation fails as a whole. Rob Hardy argues that “a poem implies an intimate encounter with another culture; a conversation; a translation not just of words, but of worlds” (Hardy, 2012). Most translations are durable creations and reflect the language of a particular place and time. That is what machines fail at. Also, poetic language is often full of puns that the machine can process and translate only literally because it has not seen such word combination before. Neologisms are another pain here since they are untranslatable and have to be handled manually. Correct translation of poetry involves more context than can ever be acquired by a machine. The crux of the matter here is that the supposed aim of translation – the transparent rendering of the full meaning of a text in another language – is unobtainable. Every translation of poetry incurs the transformation of language effects, and every target language is different in terms of prosody. Unfortunately, the operating principle of machine translation programs is such that they use only phrase-based methods: if the sentence in a poem is long, the program may not recognize it as a sentence, and, therefore, translate in chunks, and linking these chunks becomes problematic.
As it has been mentioned above, evaluation in poetic translation is crucial. In order to evaluate computer-generated poetry, translators should understand that poetry deals with the language on all its levels: semantic, rhythmic, phonetic, and syntactic. Generating and translating poetic texts, computers rely on the algorithms that were initially embedded into the translation program. In other words, the algorithms here imply search procedures, according to which the exact or fuzzy matches are found. Google differentiates between two types of algorithms: literal and natural or hybrid. Based on the literal algorithm type, the lexicon of a poem will be rendered word by word (matching one word in one language with the same word in another language); however, in order to make the poem flow more naturally, the program may recombine words (natural/hybrid algorithm type). The hybrid algorithm type is the hardest to set because it should enable the machine to “understand” the thought-provoking impact of a poem on its target reader. Also, Google’s principle of using literal or hybrid algorithms follows Eugine Nida’s concept of formal and dynamic equivalence. The latter provides a natural and easy form of expression in order to produce a similar response in the target reader (Nida, 1964). Therefore, any assessment of computer-generated poems should be largely based on the algorithm type and the concept of equivalence approach being taken. The process of evaluation in machine translation of poetry, however, should be void of a “quality” component since the latter is subjective, and computers are only seen as translation enhancement tools.
In addition, the psychology of a human translator is significant in the evaluation of computer poetry. It is impossible, however, to explain the cognitive processes in translator’s mind without combining them with the translator’s working environment, world knowledge, experience, skills, and different artifacts. Since artifacts make humans smarter, in machine translation of poetry translators largely depend on translation memory (TM) and online poetry corpora (OPC). Both TM and OPC, being parts of a cognitive process, interfere with and control translator’s mental processes in choosing equivalents and rhyming words. Another cognitive problem is the perception of the text by novices and experts. Looking at a machine translated poem, a novice translator focuses primarily on words and edits the verse on the word level; however, an expert looks at the same randomized gibberish output as a text. According to Katharina Reiss’s text classification, poems are considered operative texts because they induce behavioral responses and appeal to or persuade the reader (Reiss, 1976:15). Since computers are seen only as enhancement tools in poetic translation, and human translators are final decision-makers, it is important to know what cognitive processes underlie translator’s ultimate decisions in producing poems and poetic texts.
Overall, machine translation of poetry seems to be very possible now and it is a real option for minority languages. Richard Richens points out that “if a Georgian speaker wishes to appreciate the imagery of Welsh poetry, machine translation might be well prove to be the ideal approach” (Madsen, 2009). It is true that the output of machine translation often sounds gibberish and incoherent but this should not hinder or stop research in the field. Google’s attempts to generate and translate poems might be rewarded within the next five or ten years, just like the attempts of those computer programmers who designed the first chess programs that beat a human player in 1950s. What needs to be kept in mind, however, is that the translation of poetry is a creative art that requires from a translator to follow certain rules as well incurs subjective involvement when comparing the translated text with its original. Although many computer programs are now well-suited for the translation of simple iambic lines with one- or two-syllable words, the machines still have a long way to go in order to simulate human understanding and intuition. Translating Russian poetry, for example, where rhyme and meter are dominant elements, computer programs are always pulling between form and content, and as a result, often have to compromise semantic accuracy, word choice and order, lexical levels, etc. Of course, only the human translator can be a true mediator between the author and the reader in expressing emotions of poems; however, machine is not necessarily required to produce an exact copy of the original. It can produce a “fake” of the original: a fake poem lacks any criminality provided that it stays as close to the source text as possible and has a similar emotional impact on the target reader.
 

Works consulted

Genzel, Dmitriy. "“Poetic” Statistical Machine Translation: Rhyme and Meter." Conference on Empirical Methods in Natural Language Processing. Massachusetts, 2010. 158-166.
Hardy, Rob. Word of the day. 12 July 2012. 8 January 2013. .
Hartman, Charles. Virtual Muse. Experiments in Computer Poetry: Wesleyan, 1996.
Madsen, Mathias Winter. The Limits of Machine Translation. Thesis. Copenhagen: Department of Scandinavian Studies and Linguistics, 2009. 8 January 2013. .
Nida, Eugene A. The Theory and Practice of Translation . Brill Academic Pub, 2003.
Reiss, Katherina. Text types, translation types and translation assessment, translated by A. Chesterman, in A. Chesterman (ed.) (1989), pp. 105-15. 1977/89.
Teterin, Sergei. Cyber-Pushkin. 2006. 19 January 2013. .