AI, Decoloniality and Creative Poetry Translation

A collaborative research experiment supported by the John Fell OUP Research Fund 2024-25

Participants: Matthew Reynolds (PI), Joseph Hankinon (PDRA)
Georgina Fooks, Rowan Anderson, Alyssa Ollivier-Tabakushvili, Ola Sidorkiewicz, Karima Brooke, Yousif Qasmiyeh, Catriona Perry, Kathy Fan, Sarah Ekdawi, Tinashe Mushakavanhu (OCCT/Oxford); Kathryn Eccles, Ana Valdivia, Yui Kondo (Oxford Internet Institute); Wen-Chin Ouyang, Ret'sepile Makamane, Cosima Bruno, Yingxin Chen, Rafaa Ragebi (SOAS); Kọ́lá Túbọ̀sún (OlongoAfrica.com); Annmarie Drury (CUNY); Karen Cresci, Lisa Rose Bradford, Luciana Beroiz, Vanesa Venditti, María Eugenia Rigane (Mar del Plata); Vani Nautiyal (IIT Roorkee); Chakidipta Baruah, Sajal Pathak, Aanya Mehta (Delhi); Xinyi Wang (Tsinghua); Ellie Pavlick, Lachlan Kermode (Brown).     

The field of translation studies faces two major challenges. The first is decolonial critique. Scholars such as Makoni, Makalela and Sakai have dismantled the picture of languages as separate, standardised entities with translation operating between them and have shown it to be a colonial construction. Instead, they present language difference as a variable continuum, with translation as one of the means for negotiating it. Translation can follow the colonial blueprint, supporting the standardisation of languages and a restricted conception of equivalent meaning (this is in fact how it is used in many professional contexts). But it can also work to decolonial effect, activating the multilingual and varied language competences that most people possess. I have called this mode of translation ‘translationality’ (2016): it foregrounds process and collaboration, making visible the complexity of language differences and discovering new ways of traversing them.

The second challenge is the rapid improvement of machine translation and, now, the advent of generative AI. This has created uncertainty, with some voices asserting that human translators can be replaced by machines, while others insist that human involvement will always be required, especially for literary translation. Strikingly, the debate is happening almost entirely within the picture of translation that has been shown to be a colonial construction – i.e., a matter of success or failure in achieving equivalence of meaning between standard languages conceived as separate.

This picture fails to capture what is new about generative AI tools relation to translation: the fact that they can handle language variety in more particular ways than longer-established tools such as Google Translate. Those older tools work within the colonial regime of standardised languages, translating between something defined as eg ‘Italian’ and something defined as eg ‘Swahili’. But LLMs such as Chat-GPT, Gemini and Claude present a more nuanced landscape of language variation. You can ask them to translate into a given style, or professional register, or dialect, or the language of a historical period, or a mix of languages. You can prompt them to use a particular rhyme-scheme, or preserve a sound pattern, or use only words of one syllable. The responses to these questions are of varying fruitfulness (variable between tools, between languages involved, and between questions); but they are interesting. They show that machine translation is not necessarily the enemy of linguistic diversity and play: it has the potential to contribute to creative, decolonial translation practices.

Our collaborative experiments will investigate this potential. Focusing on poetry (typically excluded from public discourse around machine translation), we will try out how LLMs might join in creative translation practices involving a range of languages, both low-resource and high-resource, and we will investigate related political, technical and conceptual issues. Material will be shared in an online workspace, with regular meetings to talk over discoveries and ideas. The ongoing experiment will be punctuated by symposia at which we will hear from and discuss with other researchers and practitioners. Participants will keep a record of their linguistic explorations and a diary of their thoughts and feelings about them.

Sometimes – perhaps especially with low-resource languages – what is discovered will be a barrier, misrepresentation or lack: in these cases the aim will be to articulate what would be needed for creative-translational collaboration with AI to work better. Dialogues with the Language Understanding and Representation (LUNAR) Lab at Brown University and the Oxford Internet Institute will enable us to do this in an informed way, as well as to clarify what is revealed by happier discoveries. The aim is to be interested in the possibilities of LLMs for creative, decolonial translation, and for what they can show us about language and creativity, rather than to repeat familiar narratives of their limitations. Mindful of the commercial aspect of LLMs, their environmental cost, and their repercussions for the translation profession, we will give attention also to open-source and smaller models, and seek to show how our experiments might increase awareness of the importance of language diversity online and stimulate participation in creative translation practices. We hope to expand the understanding of translation in AI, and AI in translation, to include both creativity – conceived here as collaborative and distributed, more Oulipo than Romantic – and a decolonial engagement with language variety and resource inequality.

The project’s objectives are to:

(i) develop practices of creative interaction with the translation capabilities of LLMs, exploring how these capabilities can best be used in the service of a creative, decolonial literary translation practice involving many languages, including especially Global Majority languages.

(ii) provide an appraisal of the usefulness of LLMs in this kind of translation.

(iii) articulate how LLMS could be adjusted to support creative, decolonial translation practices more effectively.

(iv) work out the optimal critical practice for recording, theorizing and presenting our results, given that academic writing in standard English and fixed professional forms is itself a kind of translation that has been subjected to decolonial critique.