Duel with DeepL


Literary translator Hans-Christian Oeser on machine translation and
the translator’s voice


What experience have you had with MT tools?

I have been working as professional literary translator from English to
German for the past forty years, slowly advancing from mechanical to
electric to electronic typewriters and from a non-IBM-compatible Amstrad to a variety of PCs and laptops. In a sense, my work practice has mirrored the onward march of technological progress over the decades, albeit in a rather halting and hesitant fashion, owing to my innate conservatism. So computer-assisted translation tools were far beyond my horizon ... With the advent of the internet, my wide range of well-thumbed physical dictionaries has gradually been superseded by electronic dictionaries, although I had been quite dextrous at finding the right lemmata at a fingertip even in large volumes such as the Große Muret-Sanders (an arduously acquired skill!). Obviously, online dictionaries have the huge advantage of regularly being brought up to date, whereas printed lexica are hopelessly antiquated soon after publication.


When I did try out translation applications such as Babel Fish or Google Translate, I found them quite useful in day-to-day contexts but utterly lacking in quality and reliability when it came to literary texts. When approached by two computer linguists who wished to carry out empirical research into the benefits or otherwise of machine translation tools for literary translation, I agreed rather apprehensively to participate in two experiments, the first serving as a pilot study for a broader in-depth investigation.

In autumn 2018, I was asked to revisit F. Scott Fitzgerald’s The Beautiful and Damned, a novel I had translated into German as early as 1998. In spring 2019, I was handed an excerpt of about six pages, which I was to have processed by a machine translation program and to post-edit. The purpose of the exercise was to compare the ensuing result with my purely “human” translation of 1998 in terms of style and vocabulary and to examine “how the translator’s voice is affected in workflows involving machine translation”. It was found that my “textual voice” was somewhat diminished in my postedited work compared to its stronger manifestation in my earlier machineindependent German version. In my comments on those findings, I argued that the diminution detected by qualitative corpus analysis might also, if not more, have to do with a general impoverishment of my vocabulary which seems to have occurred over time in spite of the amassed wealth of experience as a literary translator. In other words, there was not only the opposition “man v. machine” to consider, but also that of “then v. now”.1

Which tools did you use?

My suggestion to the researchers was to use DeepL Translator, the free commercial machine translation service launched in 2017, which I judged to be more accurate than Google Translate or Bing Microsoft Translator, if by no means adequate for literary purposes. As the translation in question as well as the more comprehensive follow-up was or is to be a once-off experiment and I do not necessarily wish to proceed wholesale with machine translation, I did not purchase DeepL Pro.

I was then asked to engage in a quasisupervised translation research project on a much larger scale, the funding of which has not yet been approved. A German publishing company based in Hamburg had commissioned me to translate Christopher Isherwood’s novel The World in the Evening (published in autumn 2019). This work of 333 pages was deemed suitable for a full-length comparison of original text, machine-generated translation and post-edited version.

And thus it was that, in spring 2019, I fed the DeepL Translator window with small portions of the original (the limit for any one feed is c. 5000 characters) – a process that took, believe it or not, less than seven hours, meaning that an English-language novel of considerable length could be presented to the prospective German reader in less than one working day. But would it be a faithful and creative rendering of Christopher Isherwood in German? Would there be a recognisable translator’s voice at all?

In addition to those aesthetic questions, I fear certain repercussions not so much for not using the tool but, indeed, for using it. To my non-expert mind, there might very well be legal and contractual implications. Whose work is the finished product? The machine’s (or its producers’ and providers’), the human translator’s or both? Who can, in the end, claim copyright? Could DeepL Translator
rightfully maintain that, in spite of my post-editing efforts (which would place me in a position similar to that of a publisher’s editor and copy editor), I had appropriated “their” translation? Could the publishers contend that not I myself but rather a translation program was the originator of the German version and that therefore I should receive less pay?


Were you trained in how to use this tool?

No training was needed. It was a cakewalk.

And your overall views?

I must confess that whenever colleagues working in the domains of commercial, technical, legal or medical translation mentioned computeraided translation tools such as Trados (and that happened in the eighties!), I had to plead ignorance or expressed doubts and suspicions, being unable to abstract from my own field of literary translation with its particular and peculiar challenges of individual style, including rhythm, sound, musicality etc.

As for DeepL Translator, I would summarise my admittedly limited experience as follows. There are advantages and disadvantages in employing machine translation for works of literature. Psychologically speaking, it is comforting to have an entire novel at the ready within a few hours. It feels (and, of course, that feeling is a huge fallacy!) as though the work to be carried out might already have been accomplished – a welcome boost to your self-esteem. Also, in terms of the time spent on post-editing, as opposed to translating from scratch, it could be argued that the overall effort is somewhat less time-consuming.

On the downside, the translation, as proffered by the machine, bears no resemblance whatever to a readable – and enjoyable – human translation, deficient as the latter may be. Themachine has, as of yet, no proper sense of context, of wordplay, ambiguity, polysemy and metaphor or of rhetorical devices such as alliteration and assonance. It frequently mistranslates, using inappropriate words and phrases, seemingly chosen at random from its vast lexicon. Two examples (from another attempt at machine translation) may suffice: “Ungeöffnete Buchstaben liegen auf einem Stapel” for “Unopened letters are in a pile”, where ‘letters’ is translated in the meaning of ‘sound symbol’, not of ‘writing’. Or: “Das Bild wird gehalten, wenn Anita bei einer Landung innehält” for “The image is held when Anita pauses on a landing”, where ‘landing’ is translated as the touchdown of an aeroplane rather than as part of a staircase. These errors are easy-peasy and can be rectified without difficulty. But on a syntactical level, sentence structures often remain very “English” if they are not outright ungrammatical. Occasionally, it can be disheartening to have to disentangle the machine’s gobbledygook.

Worse still, the machine has no awareness of elegance, of beauty, of stylistic coherence (or indeed intended breach of style) and is unable to create an unmistakable “sound”, combining the original author’s personal voice with that of the translator. Its output is altogether uninspired and uninspiring. “Postediting”, in fact, entails painstaking retranslation. Hence, in the case of Christopher Isherwood’s novel there was hardly a sentence that did not have to be thoroughly revised and rebuilt.

There is another trap. When you revise a pre-existing translation, and the machine translation is a pre-existing translation and not a rough draft of your own making, you are faced with a dilemma not encountered when embarking on a fresh translation that is not machine-produced. It is a dilemma known to every editor proper: how to respect both the original author’s and the translator’s voice? You do not have to “respect” the machine’s “efforts” but you have to set to work against two backdrops at once: the original and the pre-existing translation, each posing its own difficulties. Psychologically and mentally, more often than not your creative energy is channelled along predefined paths which you might not even have known to exist and which might not at all correspond to your own writing style as it has developed over time. This might result in a constraint if not a loss of linguistic and literary competence in terms of word selection and sentence formation. Whereas, if confronted with the original alone, you are compelled to find innovative solutions of your own to each and every challenge. Your professional experience, your educational habitus, your instinctive feeling for language, your aesthetic intuition, the spontaneous inspiration of the moment will suggest words, phrases and sentence constructions utterly different from those suggested and indeed pre-empted by the machine. Your translational activity might be aided on one level but on another it is hampered and impeded.

Was your translation fee any different than if you hadn’t used an MT tool?

As no one was aware of my doings, I was paid my usual fee. However, there are justified fears among literary translators that in future publishing houses might commission book translations stipulating the use of machine translation tools in order to reduce fees and to downgrade the literary translator, who only recently has been able to secure some degree of social standing, to the role of out-of-house editor.

How do you see the future of literary translation in relation to both CAT and MT tools?

Technological progress, including the development of artificial intelligence and its offshoots, seems unstoppable, irrespective of ethical and practical considerations. There is no doubt that sooner rather than later neural machine translation tools will be further refined, incorporating ever greater quantities of text corpora and ever more subtle logarithms doing greater justice to complex grammatical features.

My savvy computer linguists seem to think that the advantages of both machine translation and human translation could be combined by “personalising” the tools on offer. By means of computer-aided text
analysis all previous works of one literary translator could be assembled as a comprehensive text corpus which would form the basis for future text generation to be used and polished by the selfsame translator.

For now, I would propose that every literary translator ought to have the possibility and the right to utilise every tool at their disposal, that is to say: not only analogue and digital dictionaries, not only translation memory and terminology management software but also online or offline translation programs of every description. On the other hand, no literary translator must be coerced into doing so or made to merely
redact machine translations, with the corresponding loss of income and status.

I, for my part, shall continue to avail of electronic tools but, being conscious of the dangers arising to my artistic autonomy, only to spot-check and not over a wide area of text. As an organised community, however, we must strive to resist and reject any attempt by publishers (some of whom are already rumoured to contemplate steps in thatdirection) to transform, as part of a cost-minimising exercise, from machineaided human translation to humanaided machine translation that which is rightly our work. We will have to be the Luddites of the humanities! For as long as publishers regard literary translation not primarily as a commercial commodity but as a product of the human intellect, the human imagination and the human spirit, for as long as they are interested in high literary quality, we may harbour some hope.


1 The findings were published in Dorothy Kenny & Marion Winters, “Machine Translation, Ethics and the Literary Translator’s Voice”, in: Translation Spaces, vol. 9, issue 1 (August 2020), pp. 123-149.

(Counterpoint, H. 4, Dezember 2020, S. 19-23)

Link zur französischsprachigen Fassung

presse : interviews

This site was designed with the
website builder. Create your website today.
Start Now