kuroalpha/reflective-translation-mt

Code and data for “Reflective Translation: Enhancing Low-Resource Machine Translation through Self-Reflection” (English→isiZulu/isiXhosa, BLEU/COMET, reflection + Self-Refine baselines). # Reflective Translation for Low-Resource Machine Translation This repository contains code and analysis for **Reflective Translation**, a reflection-guided prompting framework that improves machine translation quality for low-resource languages **without fine-tuning**. We evaluate English → isiZulu and English → isiXhosa translation using GPT-3.5 and Claude Haiku 3.5, reporting consistent BLEU and COMET improvements across reflection, prompting strategies, and confidence thresholds. This repository supports **full reproduction of all experiments, tables, and figures** in the associated paper. --- ## Paper **Reflective Translation: Enhancing Low-Resource Machine Translation through Self-Reflection** Nicholas Cheng arXiv:2601.19871 (2026) 📄 Reflective Translation: Imp… ```bibtex @article{cheng2026reflective, title={Reflective Translation: Enhancing Low-Resource Machine Translation through Self-Reflection}, author={Cheng, Nicholas}, journal={arXiv preprint arXiv:2601.19871}, year={2026} } ``` ## Overview Reflective Translation introduces a lightweight, inference-time framework in which a language model: 1. Produces an initial translation 2. Generates a structured self-critique identifying translation errors 3. Produces a revised translation guided by masked reflective feedback This approach improves semantic fidelity and robustness in low-resource settings without additional training data or parameter updates. ## Datasets All datasets are loaded programmatically from HuggingFace at runtime. OPUS-100 (English–isiZulu, English–isiXhosa) huggingface.co NTREX-African (isiZulu, isiXhosa test sets) ntrex_african Raw parallel corpora are not redistributed in this repository. ## Repository Structure ``` ├── src/reflective_translation/ # Core package (prompts, metrics, pipeline) ├── scripts/ # Experiment and figure generation scripts ├── outputs/ │ ├── …