TLDR: Researchers have developed a new method to significantly improve machine translation for low-resource languages like Wayuunaiki. Their approach combines supervised fine-tuning with reinforcement learning, allowing large language models to effectively use an external bilingual dictionary. This dictionary-guided strategy, particularly with the Qwen2.5-0.5B-Instruct model, achieved notable improvements in Spanish-to-Wayuunaiki translation quality, demonstrating the power of integrating external tools and reinforcement learning for underrepresented languages. The study shows up to a +3.37 BLEU improvement and an 18% relative gain over baselines.
The field of natural language processing (NLP) has seen tremendous advancements, yet these breakthroughs have largely overlooked languages with limited digital resources, especially Indigenous languages. This scarcity of high-quality parallel data and the predominance of oral traditions mean that even the most advanced generative AI systems struggle to produce reliable translations for these languages. A recent study highlighted that AI responses in Indigenous languages are often less accurate, shorter, and lack fluency compared to high-resource languages.
Addressing this critical gap, a new research paper titled “Improving Low-Resource Translation with Dictionary-Guided Fine-Tuning and RL: A Spanish-to-Wayuunaiki Study” introduces a novel approach to enhance machine translation for low-resource languages. Focusing specifically on the Spanish-to-Wayuunaiki language pair, this study proposes integrating an external dictionary tool and training models end-to-end using reinforcement learning (RL), in addition to traditional supervised fine-tuning (SFT).
The Challenge of Low-Resource Languages
Wayuunaiki, an Arawakan language spoken by approximately 420,000 people in Colombia and Venezuela, serves as a prime example of a low-resource language in the NLP domain. Despite its relatively large number of speakers, it lacks extensive datasets and applications. Previous efforts to build Wayuunaiki–Spanish translation systems have faced limitations due to data scarcity and narrow topical coverage, resulting in modest performance.
A Novel Hybrid Approach
The researchers frame translation as a tool-augmented decision-making problem, where the language model can selectively consult a bilingual dictionary during the translation process. Their method combines supervised instruction tuning with Guided Reward Policy Optimization (GRPO), a reinforcement learning technique. This allows the model to learn not only how to translate but also when and how to effectively use the dictionary tool.
The training pipeline involves two main phases:
First, a large language model, specifically Qwen2.5-0.5B-Instruct, undergoes supervised fine-tuning. In this stage, the model is taught to produce structured outputs and to properly invoke the dictionary tool. This is achieved by providing examples of Spanish-Wayuunaiki translation pairs, augmented with artificial examples of dictionary lookups. Even though these artificial lookups might not always be directly useful for the translation, they help the model acquire the structured habit of using the tool.
Second, the fine-tuned model proceeds to a reinforcement learning stage using the GRPO framework. Here, the model generates multiple candidate translations for a given input. Each generated output is then evaluated using BLEU similarity scores, which serve as a reward signal. This reward guides the model to iteratively refine its translation strategy, improving overall performance and learning to use the dictionary tool more effectively. The dictionary used for this purpose was a filtered version of Rafael José Negrette Amaya’s bilingual Wayuunaiki–Spanish dictionary, containing approximately 29,000 entries.
Key Findings and Improvements
The preliminary results are highly promising. The tool-augmented models achieved up to a +3.37 BLEU improvement over previous work and an 18% relative gain compared to a supervised baseline without dictionary access on the Spanish–Wayuunaiki test set from the AmericasNLP 2025 Shared Task. The study found that performance consistently improved at each training stage, with SFT providing the largest gain and RL delivering an additional 11% improvement.
Crucially, the external dictionary tool provided a relative performance boost of approximately 6% in both the SFT and SFT+RL stages. The best-performing model, Qwen-0.5B+SFT+RL, made the most extensive use of the dictionary, averaging 3.94 calls per sample (close to the maximum of 4 allowed) with a 95% success rate in querying the dictionary.
The research also explored different model architectures, including LLaMA3.2 and larger Qwen models (Qwen2.5-7B). Instruction-tuned models consistently benefited from both SFT and RL with tool access, with larger models generally achieving better results. Qwen2.5-7B+SFT+RL achieved the highest average BLEU score of 4.45, effectively doubling its base performance.
An interesting insight was that the models did not simply copy translations from the dictionary. Statistical analysis showed that the models produced translations significantly better than merely selecting the best dictionary result, suggesting they effectively refine and enhance suggestions using their learned language knowledge.
Also Read:
- Boosting Grammar Correction Accuracy with LLMs and Rule-Based Reinforcement Learning
- Advancing Mathematical Autoformalization with Unlabeled Data Using FormaRL
Limitations and Future Directions
Despite the success, the study acknowledges several limitations. The effectiveness of the dictionary tool was constrained by its limited coverage and quality, as only a small percentage of unique Spanish words in the test set appeared in the dictionary with matching Wayuunaiki translations. The choice of reward signal in the RL stage was also critical; BLEU proved effective, while a character-level metric led to performance degradation.
Computational constraints limited the scale and duration of training, and the absence of native Wayuunaiki speakers for qualitative analysis means that a thorough evaluation of fluency and cultural appropriateness is still pending. However, the methodology is broadly applicable to other low-resource languages, especially non-agglutinative ones where words are easier to translate independently.
This research highlights the significant promise of combining large language models with external tools and reinforcement learning to improve translation quality in low-resource language settings. For more details, you can read the full research paper here.
