University of Melbourne WMT 2026 submission distills Gemini output to fill Tok Pisin, Bislama and Solomon Pijin translation gaps

Machine translation for Pacific creoles has long been limited by how little text exists to train on, and how unevenly what does exist is spread across domains like news, conversation or government documents. The University of Melbourne team’s WMT 2026 submission tackles that imbalance directly: after pre-training on a large but domain-imbalanced collection, they fine-tune on a mix deliberately balanced across domains, then patch the domains their raw data still lacked with text distilled from Gemini.

The paper frames this as one tool among several, alongside LLM-assisted respelling and alignment of existing text, and back-translation, which generates parallel data by translating monolingual text and then translating it back. Distillation from Gemini is reserved specifically for domains where no source text existed at all in the training set, meaning the model is not just cleaning up scarce data but manufacturing coverage that otherwise would not exist.

The authors evaluate their models on Bouquet, an established multilingual benchmark, and on a new test set built from spoken language transcripts, a domain that written corpora tend to miss entirely. Measured against human-original references, the resulting models beat open baseline models by a margin the paper reports directly.

The numbers

The paper reports the models beat open baselines by 3+ chrF++ points in all translation directions tested against human-original references, a metric that scores translation quality by comparing character sequences rather than whole words, which tends to reward languages with irregular or improvised spelling such as these creoles.

The team names its next steps as building human-translated test sets for Solomon Pijin and Bislama specifically, since the current evaluation leans on Bouquet and the transcript set rather than dedicated human translations for those two languages, and distilling their best models into much smaller ones that keep broad domain coverage.

For a language with a training corpus this thin, relying on a commercial model like Gemini to invent missing domain text is a practical shortcut, but it also means the resulting system’s blind spots are partly inherited from Gemini’s own blind spots in Tok Pisin, Bislama and Solomon Pijin, a chain of dependency the paper does not measure directly.

Note: I omitted the own-ledger citation since the topic “research” only accepts paper, newsroom, or docs sources, and I did not use any ledger-derived figures in the prose, so the citation requirement doesn’t apply here.