Iterative Translation Refinement with Large Language Models

Anonymous

TL;DR: We experiment with iterative translation refinement with a large language model and observe improved fluency and naturalness.

Abstract: This paper argues that benefiting from vast pre-training data, large language models offer a means to improve translation fluency. We propose iterative refinement prompting, which is infeasible for conventional encoder-decoder models. In our experiments, multi-pass querying reduces string-based metric scores, but neural metrics suggest comparable or improved quality. Human evaluations indicate better fluency and naturalness compared to initial translations and even human references, all while maintaining quality. Ablation studies underscore the importance of anchoring the refinement to the source and a reasonable seed translation for quality considerations. We also discuss the challenges in evaluation and relation to human performance and translationese.

Paper Type: short

Research Area: Machine Translation

Contribution Types: NLP engineering experiment, Position papers

Languages Studied: English, German, Chinese, French, Japanese, Ukrainian, Czech