Scaling Zero-shot Machine Translation on Low-Resource Languages with Linguistic Grammars

Most of the world’s languages lack parallel corpora, monolingual web data, or sufficient representation in multilingual LLM pretraining. However, many are documented in descriptive grammars containing exhaustive syntactic information, interlinear glossed examples and translations. Recent work has shown that large language models can leverage grammars in-context for zero-shot translation and typological classification, but it remains unclear whether grammars alone can serve as the primary supervision for training machine translation systems. We introduce a scalable grammar-centered framework for machine translation, converting descriptive grammars into structured LLM-readable context by extracting example sentences, glosses, and leveraging available external typological metadata. Across five typologically diverse low-resource languages, we present structure-aware regimes that incorporate gloss-level information and grammar conditioning to enable in-context learning for machine translation. We further analyze how translation quality scales with the number of grammar-derived examples – for Georgian (49.01 ChF), Chamorro (53.32 ChF), Basque (63.90 ChF), Igbo (52.83), and Korean (37.09 ChF) languages. Our results show that (1) grammars alone can support non-trivial translation performance when nothing else is available, (2) incorporating typological metadata does not show consistent generalization over flat sentence pairs, and (3) sentence-level examples matter most: extracted parallel examples consistently improve translation quality relative to zero-shot and typology-only prompting. This work reframes descriptive grammars as computational resources rather than static references, offering a practical pathway toward scaling machine translation to languages that have no corpus data. The code for the project will be publicly available at: github.com