The Untranslatable as Infrastructure
What machine translation flattens
This event is part of a demonstration of what Lera offers for running knowledge events. Its content is written with care and draws on real sources, but the conference has not taken place, and apart from the convener its participants are fictional.
In the Dictionary of Untranslatables, Barbara Cassin does not treat the untranslatable as what is not translated, but as what we never stop translating, and never stop failing to translate. The untranslatable is not a gap; it is a site of work. Every time a translator hesitates over a word, they reveal a piece of knowledge that has no ready place in the target language.
The history of Japanese offers a clear example of this work. Akira Yanabu (1982) showed that words such as shakai for society and jiyū for freedom were coined, or torn from their earlier meanings, in the Meiji era to carry European concepts. For decades these words hung between two worlds, neither fully European nor fully Japanese. It was this suspension that gave life to thought: speakers of Japanese had to ask what freedom really meant, because the word did not yet know the answer in advance.
Machine translation cannot bear this suspension. A model that chooses the most probable rendering turns an open word into a closed one, and each time it does so it pulls the text closer to the average. Until recently, many systems also translated between two non-English languages by way of English; a Persian text passed through the filter of English concepts before it became Japanese. Emily Bender and her colleagues (2021) warned that large models reproduce the dominant viewpoints in their data. In translation, the dominant viewpoint takes the form of word choice and is almost invisible.
Lawrence Venuti (1995) complained of the ‘invisibility of the translator’: we consider a translation good when it does not read like a translation. Machine translation has carried this ideal to its limit. It gives us fluent text with no trace of hesitation, doubt or choice. For a reader who wants to know, that fluency is misleading.
My proposal is to make the untranslatable part of the infrastructure of knowledge. That means a system that keeps marked words marked, gives each of them an entry in the glossary around the text, treats the translator’s note as part of the work rather than its margin, and keeps the original always one step away from the reader. A machine can propose a draft; but where two languages are in conflict, the decision belongs to a person whose name stands under the translation.
Questions for discussion
- How can a knowledge system mark a word as ‘untranslatable’ without making reading harder?
- Does fluent machine translation reduce the human translator from decision-maker to post-editor?
- Should the translator’s note belong to the body of the work or to the orbit around it?
References
Cassin, Barbara, ed. 2014. Dictionary of Untranslatables: A Philosophical Lexicon. Translated by Steven Rendall et al.; translation edited by Emily Apter, Jacques Lezra and Michael Wood. Princeton: Princeton University Press.
Yanabu, Akira. 1982. Hon’yakugo seiritsu jijō [How Translation Words Came to Be]. Tokyo: Iwanami Shoten.
Venuti, Lawrence. 1995. The Translator’s Invisibility: A History of Translation. London: Routledge.
Bender, Emily M., Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–623.