Adaptive Translation Engine for Low Resource Languages

The Indo-Pacific region boasts immense linguistic diversity. However, many of these languages lack support from commercial translation services, with NLP advances primarily focused on languages from affluent societies. Notably, most commercial services perform high-level analyses post-translation, as opposed to the original language. In collaboration with USCÆs Information Sciences Institution (ISI), the goal is to use and extend the Massively Multilingual Neural Machine Translation (NMT) system to cater to low-resource languages in the Indo-Pacific region. This NMT system, developed previously by the ISI team with DARPA funding, offers translation capabilities for over six hundred languages into English. In this project we propose to: 1) Deploy the NMT model for many-to-English multilingual translation for low-resource languages; 2) Fine-tune the NMT model for two specific languages: Tok Pisin (TPI) and Buginese (BUG) using limited training data. Improvements will be assessed by comparing BLEU scores; 3) Develop a web-based user interface to allow users to adapt the parent model for any low-resource language, showcasing its stackability. The research team comprises leading figures in AI and NLP, with vast experience in government and commercial projects. InferLink, the research firm, has a strong track record of turning SBIR research into commercial technology, having spawned several startups from government-funded tech in recent years. Our team members from ISI have a great deal of published research, as well as commercial and open source products in the areas related to this proposal, including machine translation, multilingualism, and NLP in general.