AMTA - AI and the Future of Low-Resource Languages

Announcement, Colab. At AMTA 2026 me and Vera Senderowicz Guerra gave an overview of the state of the art of LLM usage to translate English into a selection of languages, from low-frequency ones such as Yoruba, Basque and Quechua, to lower-frequency variants of popular languages, such as the regional Latina American dialects of Spanish. Each language and dialect has its own characteristics that brings different challenges to LLM translation pipelines. During the presentation we demoed a few aproaches that can be adopted to try improving the fluency of the LLM of your choice in your target language, from raising the reasoning budget to translation reviews, full-blown finetunes, backtranslations, and glossary injection, highlighting the specific issue that each method addresses most effectively. All code is made available in Colab here, and can be run with a T4 GPU environment and an Openrouter API key.