Accurate, Cost-Effective Translation System for Low Resource Languages

Barron Associates proposes a Phase I program that will develop the Accurate, Cost-Effective Translation System for Low Resource Languages (ACE-Translation) to address the objectives of SBIR topic A234-021: Machine Translation for Indo-Pacific Low Resource Languages.Ā ĀDuring Phase I, the team will develop methods to automate collection of training data, explore optimal structures for training datasets given constrained resources, implement machine translation systems for representative languages, develop novel quality assurance methods for the translation systems, and quantify performance of the systems using standard benchmarks.Ā The proposed approach builds on proven technologies for translation of low-resource languages previously developed by team members, as well as experience with automated data collection and assurance of critical software systems.Ā ĀThe team will build on open-sourced pre-trained models like mBART, DeltaLM, or NLLB, adapting these pre-trained models to new languages.Ā The team will employ hierarchical adapter units, which are injected between the layers of the pre-trained model and can be trained on a new language, domain, or task. There are two advantages in fine-tuning only these adapter components.Ā Since they consist of only a small number of parameters, they can be adequately trained with a small number of training examples. In addition, as the pre-trained model remains invariant, they render catastrophic forgetting a non-issue. The proposed work will improve existing processes for using these adapter units by incorporating additional phylogenetic linguistic information in their organization, leveraging results of prior linguistic studies that provided the needed information in the form of phylogeny trees.Ā The team will develop innovative automated data collection methods to enable low-cost collection of data to support implementation of translation systems for nearly any low-resource language.Ā The proposed work will also include development of strategies for optimally selecting input data for system training, including targeting data to provide enhanced accuracy in domains of interest to the military.Ā Recognizing the dangers of ōhallucinationsö (grossly inaccurate but often articulately phrased outputs) that can be produced by even the best modern language models, ACE-Translation will employ run-time assurance (RTA) to achieve high reliability.Ā Barron will leverage its expertise in RTA methods for safety critical software systems to enable ACE-Translation to detect and automatically remediate degraded performance, falling back when needed to a robust translation mode that prioritizes accuracy over fluency.Ā Phase I research will position the team for subsequent development of operational translation systems for a larger set of languages, integration of these into workflows that support Army missions, and continued enhancement of the quality of the core translation systems.