Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reflective Translation: Improving Low-Resource Machine Translation via Structured Self-Reflection

Nicholas Cheng

TL;DR

Low-resource MT on isiZulu and isiXhosa suffers from limited parallel data. The authors propose Reflective Translation, a prompt-based inference-time correction that generates an initial translation, a structured critique, and a revised translation guided by the critique, with masking to prevent verbatim copying. Evaluations on OPUS-100 and NTREX-African across two LLMs show consistent improvements in BLEU and COMET for second-pass translations, with gains emphasizing semantic fidelity. The work provides a reflection-augmented dataset and demonstrates a practical, model-agnostic method that can enhance translation quality without fine-tuning, offering a scalable path for improving low-resource MT.

Abstract

Low-resource languages such as isiZulu and isiXhosa face persistent challenges in machine translation due to limited parallel data and linguistic resources. Recent advances in large language models suggest that self-reflection, prompting a model to critique and revise its own outputs, can improve reasoning quality and factual consistency. Building on this idea, this paper introduces Reflective Translation, a prompt-based framework in which a model generates an initial translation, produces a structured self-critique, and then uses this reflection to generate a refined translation. The approach is evaluated on English-isiZulu and English-isiXhosa translation using OPUS-100 and NTREX-African, across multiple prompting strategies and confidence thresholds. Results show consistent improvements in both BLEU and COMET scores between first- and second-pass translations, with average gains of up to +0.22 BLEU and +0.18 COMET. Statistical significance testing using paired nonparametric tests confirms that these improvements are robust. The proposed method is model-agnostic, requires no fine-tuning, and introduces a reflection-augmented dataset that can support future supervised or analysis-driven work. These findings demonstrate that structured self-reflection is a practical and effective mechanism for improving translation quality in low-resource settings.

Reflective Translation: Improving Low-Resource Machine Translation via Structured Self-Reflection

TL;DR

Low-resource MT on isiZulu and isiXhosa suffers from limited parallel data. The authors propose Reflective Translation, a prompt-based inference-time correction that generates an initial translation, a structured critique, and a revised translation guided by the critique, with masking to prevent verbatim copying. Evaluations on OPUS-100 and NTREX-African across two LLMs show consistent improvements in BLEU and COMET for second-pass translations, with gains emphasizing semantic fidelity. The work provides a reflection-augmented dataset and demonstrates a practical, model-agnostic method that can enhance translation quality without fine-tuning, offering a scalable path for improving low-resource MT.

Abstract

Low-resource languages such as isiZulu and isiXhosa face persistent challenges in machine translation due to limited parallel data and linguistic resources. Recent advances in large language models suggest that self-reflection, prompting a model to critique and revise its own outputs, can improve reasoning quality and factual consistency. Building on this idea, this paper introduces Reflective Translation, a prompt-based framework in which a model generates an initial translation, produces a structured self-critique, and then uses this reflection to generate a refined translation. The approach is evaluated on English-isiZulu and English-isiXhosa translation using OPUS-100 and NTREX-African, across multiple prompting strategies and confidence thresholds. Results show consistent improvements in both BLEU and COMET scores between first- and second-pass translations, with average gains of up to +0.22 BLEU and +0.18 COMET. Statistical significance testing using paired nonparametric tests confirms that these improvements are robust. The proposed method is model-agnostic, requires no fine-tuning, and introduces a reflection-augmented dataset that can support future supervised or analysis-driven work. These findings demonstrate that structured self-reflection is a practical and effective mechanism for improving translation quality in low-resource settings.
Paper Structure (21 sections, 2 equations, 7 figures, 1 table)

This paper contains 21 sections, 2 equations, 7 figures, 1 table.

Table of Contents

  1. Introduction
  2. Contributions.
  3. Methods
  4. Reflective Translation
  5. Masking to Reduce Copying
  6. Datasets
  7. Experimental Setup and Evaluation
  8. Baselines and Prompting Strategies
  9. Metrics
  10. Results
  11. Data Analysis
  12. First- vs. Second-Pass Improvements
  13. Effect of Confidence Thresholding
  14. Statistical Significance Testing
  15. Discussion
  16. ...and 6 more sections

Figures (7)

  • Figure 1: Overview of the reflective translation framework: first-pass translation $\rightarrow$ structured reflection (with masking) $\rightarrow$ second-pass translation.
  • Figure 2: First vs. second attempt translation quality by prompting strategy. (a) COMET. (b) BLEU.
  • Figure 3: Threshold ablation: average BLEU and COMET before vs. after reflection across thresholds.
  • Figure 4: Baseline translation prompt for the first attempt.
  • Figure 5: Baseline second attempt with reflection.
  • ...and 2 more figures