LuxMT Technical Report
arXiv:2602.15506v1 Announce Type: new Abstract: We introduce LuxMT, a machine translation system based on Gemma 3 27B and fine-tuned for translation from Luxembourgish (LB) into French (FR) and English (EN). To assess translation performance, we construct a novel benchmark covering LB-FR, LB-EN, and LB-FR using human-translated data from Luci, a tourist magazine about Luxembourg. Training data stems from LuxAlign, a parallel corpus of multilingual Luxembourgish news articles, and LB parliamentary transcripts augmented with Google Translate. We filter the data using LuxEmbedder, LB sentence embeddings, to remove low-equivalence segment-pairs. Overall, LuxMT's results suggest strong improvements over the Gemma 3 baseline, even for translating LB to German (DE), despite the training data not containing any DE. We also explore LuxEmbedder's potential to be used as a quality estimation metric and find strong correlations with other reference-based metrics. However, we call for further rese
arXiv:2602.15506v1 Announce Type: new Abstract: We introduce LuxMT, a machine translation system based on Gemma 3 27B and fine-tuned for translation from Luxembourgish (LB) into French (FR) and English (EN). To assess translation performance, we construct a novel benchmark covering LB-FR, LB-EN, and LB-FR using human-translated data from Luci, a tourist magazine about Luxembourg. Training data stems from LuxAlign, a parallel corpus of multilingual Luxembourgish news articles, and LB parliamentary transcripts augmented with Google Translate. We filter the data using LuxEmbedder, LB sentence embeddings, to remove low-equivalence segment-pairs. Overall, LuxMT's results suggest strong improvements over the Gemma 3 baseline, even for translating LB to German (DE), despite the training data not containing any DE. We also explore LuxEmbedder's potential to be used as a quality estimation metric and find strong correlations with other reference-based metrics. However, we call for further research to fully assess the metric's utility and advise using it with caution.
Executive Summary
The LuxMT Technical Report introduces a machine translation system fine-tuned for translating Luxembourgish (LB) into French (FR) and English (EN) using the Gemma 3 27B model. The study constructs a novel benchmark using human-translated data from Luci magazine and a parallel corpus of multilingual Luxembourgish news articles and parliamentary transcripts. The results show significant improvements over the baseline model, even for LB to German (DE) translations, despite no DE training data. The report also explores the potential of LuxEmbedder as a quality estimation metric, finding strong correlations with other metrics but calling for further research. The study highlights the potential of fine-tuned models and embeddings in enhancing translation quality and quality estimation.
Key Points
- ▸ Introduction of LuxMT, a machine translation system fine-tuned for LB-FR and LB-EN translations.
- ▸ Construction of a novel benchmark using human-translated data from Luci magazine.
- ▸ Significant improvements over the Gemma 3 baseline, including for LB-DE translations.
- ▸ Exploration of LuxEmbedder as a quality estimation metric with strong correlations to other metrics.
- ▸ Call for further research to fully assess LuxEmbedder's utility and advice to use it with caution.
Merits
Innovative Approach
The study introduces a novel machine translation system and benchmark, leveraging fine-tuning and embeddings to enhance translation quality.
Strong Performance
LuxMT demonstrates significant improvements over the baseline model, even for languages not included in the training data.
Quality Estimation Insights
The exploration of LuxEmbedder as a quality estimation metric provides valuable insights into its potential utility.
Demerits
Limited Training Data
The reliance on augmented data from Google Translate may introduce biases and inaccuracies.
Need for Further Research
The study acknowledges the need for further research to fully assess the utility of LuxEmbedder as a quality estimation metric.
Caution Advised
The report advises using LuxEmbedder with caution, indicating potential limitations in its current form.
Expert Commentary
The LuxMT Technical Report presents a significant advancement in the field of machine translation, particularly for low-resource languages like Luxembourgish. The fine-tuning of the Gemma 3 27B model and the construction of a novel benchmark using human-translated data from Luci magazine demonstrate a rigorous and innovative approach. The study's findings, particularly the strong performance in LB-DE translations despite the absence of DE training data, suggest the potential for transfer learning and the effectiveness of embeddings in enhancing translation quality. The exploration of LuxEmbedder as a quality estimation metric is particularly noteworthy, as it provides valuable insights into the potential utility of embeddings in this context. However, the study acknowledges the need for further research to fully assess the metric's utility and advises using it with caution. This caution is well-founded, as the reliance on augmented data from Google Translate may introduce biases and inaccuracies. Overall, the study contributes significantly to the field of machine translation and quality estimation, highlighting both the opportunities and challenges in translating low-resource languages.
Recommendations
- ✓ Further research should be conducted to fully assess the utility of LuxEmbedder as a quality estimation metric.
- ✓ Investment in high-quality, human-translated data for low-resource languages should be prioritized to enhance the accuracy and reliability of machine translation systems.
