Adapting a pretrained machine translation model to new terminology typically means gathering parallel data and fine-tuning the model again. For low-resource languages, that process is often prohibitively expensive. A new arXiv paper, GLOSS-MT, asks whether controllability can instead be added through an editable dictionary.
The paper's abstract is cut off before describing the method's evaluation, so the approach is presented as a proposal rather than a demonstrated result. The motivation is clear: avoid the cost of fine-tuning by letting users edit a dictionary that steers the model's output.
If the method works, it could make terminology control more practical for low-resource translation. But until the full paper is available, the effectiveness remains unverified.