Humor and Translation: Artificial Intelligence and Literary Translation
By Mark Herman
Even at its current stage of development, artificial intelligence (AI) can rapidly produce massive amounts of translated material, including material falling into the broad category of “literature.” AI can deal with metaphors and rhymes and produce text in the style of specified authors. Compared to older machine translation programs, AI makes far fewer gross errors, such as translating “Out of sight, out of mind” into “Invisible, insane.” But can AI, in the absence of human intervention, now actually produce anything to rival the work of a competent human translator? The answer, as determined after reading two articles in a recent issue of Translation Review, is no.
The first article is a review by Kim Horner of Computer-Assisted Literary Translation1, a book of essays about current work in the field. These essays generally state that: 1) literary translators should consider computer-assisted literary translation (CALT) an aid, rather than a threat, 2) human intervention will probably always be needed at some point in a literary translation, and 3) translators who dismiss CALT risk becoming obsolete. According to Andrew Rothwell, one of the book’s editors, “recent developments suggest that…blanket rejection is no longer appropriate (if it ever was)” (107 in the book, Horner 96). These developments include neural machine translation, which, developers claim, allows a machine to learn in much the same way a brain does (Horner 96).
The article does point out potential problems. Obviously, simply accepting what a machine prints out as good enough will impoverish literature. However, such impoverishment may also occur even in works that are post-edited by humans because, as one study discovered, translators tended to accept the AI’s suggestions, which led to a homogenization of translations. Also, predictably, the human translators’ engagement with the original text was more superficial than that of human translators who worked without non-human assistance.
Another study found that a human translation of a certain novel was much better than a human post-edited AI translation because the human-only translation had more of the non-standard linguistic features also present in the original novel: “subject-verb inversion, the preservation of lexical richness, higher register when appropriate, and an attempt to replicate…spoken language” (Horner 97).
On the other hand, as another essay maintains, by almost instantly making a reasonably good first draft, AI can free the human translator to “concentrate on the more stimulating higher-level challenges” (122 in the book, Horner 97). Furthermore, improved AI translation will improve the quality of translations by nonprofessionals, such as translations of internet texts.
An interesting section of the book deals with computer aids for literary translators that do not actually translate, such as a program that quantifies “word frequencies, correlations, and other…data” of the original to better help the human translator reproduce the original author’s style (Horner 98).
The other article is “A Bard is Born: A New Era of Poetry Translation by ChatGPT-4.” Its author, Xingzhong Guan, a professor at the School of Translation and Interpreting in Beijing, compared the ability to translate poetry of the AI program ChatGPT-4 with those of two older translation programs and with those of human translators.2 Most of the article is concerned with the classic Chinese poem “Spring River Flower Moon Night,” the best-known poem by Zhang Ruoxu (c. 670 – c. 730), a poet of the early Tang dynasty. Much less space is given to the AI translations of two other poems, “Szabadság, szerelem!” by Sándor Petőfi (1823 – 49), considered to be the national poet of Hungary, and “Das aesthetische Wiesel” by the German poet Christian Morgenstern (1871 – 1914).
Xingzhong Guan showed that the two older translation programs, Bing Translate and Google Translate, came up with nonsense when attempting to translate the Chinese poem. For that poem, AI compared well with the human translators. Unfortunately, that is because the human translations are not very good. Here is the title and first four lines of the AI translation:
A Spring River Flows Silently at Night
Amidst springtime tides by the sea so wide,
With a moon that emerges on the ocean’s tide.
Endless ripples follow, a thousand miles they roam,
Where’s a spring river without a moon to call home? (31)
Ignoring the erroneous punctuation at the ends of the second and third lines, the translated poem itself is not very good. The meaning is generally correct but it is obvious that AI only considered the rhyme words to be important. In order to rhyme, there are non-fluent inversions at the ends of the first and third lines; the insertion of the word “so” in the first line does not help. Also, the meaning of the fourth line is strained.
Here is the AI translation of the Hungarian poem:
Freedom, love, my soul’s desire,
Both ignite my inner fire.
For my love, I’d give my all,
For freedom’s call, my love I’d recall. (35)
The Hungarian poem has repeated sounds. That is probably the reason for the three “all’s,” but they do not produce an analogous effect in English. The last line is metrically awkward. The AI needs to be instructed, if possible, on how to intimate rhymed, metered form in rhyme-poor English without slavishly following the form of the original poem.
For the Morgenstern poem, Xingzhong Guan asked AI “to preserve the whimsical and humorous elements of the original text in its translation” (36).
Here is Morgenstern’s poem, set up as originally published, together with a literal English translation:
Here is the AI translation, which does not follow the form of the original, together with a translation by Ronnie Apter and me:
Xingzhong Guan claims that the AI “translation is colloquial, convivial, and playful. It captures the essence of the weasel’s unexpected behavior and the whimsical explanation provided by the mooncalf, inviting readers to embrace the absurdity and joy of the poetic imagination.” (36)
I disagree. Among other things, the words “near” and “dear,” though they have the same “-ear” ending as four rhyme words in the original German, are superfluous words inserted just for the sake of a rhyme. In general, the AI version has too many words that drown rather than point up the humor. Finally, and perhaps most important, the AI program did not realize that, in the original poem, the whole point of the first three rhyming lines is that the second and third lines rhyme with “weasel.” Their literal meaning is unimportant. They just have to sort of make sense, be absurd, and rhyme with “weasel.” Morgenstern’s humor depends on peculiarities of the German language, and translations can only be successful if they use corresponding peculiarities of the target language.
In general, this article suffers from the fact that Xingzhong Guan’s standards for poetry, or at least for poetry in English, are too low. None of the translations he praises, whether done by humans or by machines, are actually good. He does show that AI translation is susceptible to instruction and that different instructions can produce different translations. Good instructions to produce better translations could include such things as whether or not to be literal, whether or not to rhyme, and whether or not to follow the form of the original. However, such instructions, even if they prove to be effective, will need to be given not only poem by poem, but often line by line, and by a person with poetic sensibility. All in all, Xingzhong Guan’s article demonstrates the main point of Kim Horner’s review: AI cannot now, when dealing with literary translation, do more than assist a human being.
Notes
1. Horner, Kim. “Review of Computer-Assisted Literary Translation, Andrew Rothwell, Andy Way, and Roy Youdale (eds.), (Routledge, 2024),” Translation Review (Vol. 120, No. 1, 2024), 96-98.
2. Guan, Xingzhong. “A Bard Is Born: A New Era of Poetry Translation by ChatGPT-4,” Translation Review (Vol. 120, No. 1, 2024), 23-39.
3. Morgenstern, Christian. Galgenlieder, Palmström, Palma Kunkel, Der Gingganz (Philipp Reklam jun., 1978), 18.
Submit items for future columns via e-mail to mnh18@caa.columbia.edu. Discussions of the translation of humor and examples thereof are preferred, but humorous anecdotes about translators, translations, and mistranslations are also welcome. Include copyright information and permission if relevant.
