Can Translations Be Less Translated? Leveraging Prompts to Mitigate Translationese

28 November 2023 · Christoph Purschke · 2 minute read · #conversations

CuCo Lab conversations | Maria Kunilovskaya (Saarland University)

INFORMATION

Translated texts are known to exhibit translationese properties that make them distinct from comparable non-translations in the target language along a number of linguistic parameters. As a result, translations can be more difficult to understand for the target audiences, incurring additional processing costs. The difference between translations and non-translations is known to have an effect on the evaluation of machine translation engines. Translations might misrepresent the target language in contrastive studies. This project applies Large Language Model (LLM) prompting, an emerging NLP method exploiting the capacities of generative AI to solve linguistic problems, to the task of debiasing translationese. We use a set of linguistically motivated translationese indicators in a series of prompt design experiments to arrive at a translation variant that is less translated than the input target text. In effect, LLM is prompted to post-edit an existing human translation segment-by-segment following an instruction customised to each individual input segment. The task of debiasing translationese (detecting and eradicating the translationese features) can be useful for producing a more homogeneous output in the target language that exhibits less deviations from the expected target language norm.

This study contributes to the growing bulk of work that demonstrates the benefits of using LLM for text adaptation such as simplification, style transfer, or translation (post-)editing.

The prompt-design part of this project explores human-machine interaction options, rising to the challenge of combining the explanatory power of knowledge-rich features rooted in contrastive analysis and translation theory with the computational capabilities of modern language models. The experiments are based on a bidirectional German-English sample from the parallel Europarl data.

SPEAKER

  • Dr. Maria Kunilovskaya, postdoctoral researcher at Saarland University

COORDINATES