Accurate, Cost-Effective Translation System for Low Resource Languages

The Phase I research effort demonstrated the feasibility of ACE-Translation (Accurate, Cost-Effective Translation), a system for efficiently deploying high-performance machine translation for low-resource languages.   The Phase I work demonstrated: (1) high-performing neural Machine Translation (MT) systems could be implemented using very small quantities of parallel data, (2) the effectiveness of novel approaches for exploiting grammar information to improve MT system performance in extremely-low-resource settings, and (3) the feasibility of building robust quality assurance systems for MT in low-resource settings.  A quality assurance system was implemented leveraging a Run-Time Assurance (RTA) framework previously developed by the team for safety-critical control systems, and a complete MT system with quality assurance monitoring was implemented and demonstrated for the Ivatan language.  Building on this success, the team proposes a Phase II program to develop and demonstrate a comprehensive set of tools to support efficient deployment of assured MT systems for low-resource languages, including deployment of new systems for languages not previously supported, and performance optimization for specific domains and/or for specific downstream tasks.  The ACE-Translation design emphasizes assurance of the quality and correctness of outputs.  One aspect of enhancing quality is leveraging as much of the available information as possible to produce high quality translations for a language of interest. In high-resource settings, where abundant parallel data exists, it is not as critical to make use of other types of data. In low-resource settings, significant gains can be realized by using additional types of data.   Grammar descriptions exist for approximately 4,300 of the world’s roughly 7,000 languages, making them a rich source of information with good coverage of low-resource languages.  Expanding capabilities for leveraging this data is an important aspect of Phase II. A second key aspect of enhancing quality is innovative use of RTA methods. In ACE-Translation, the RTA system uses three parallel and independent translation branches along with a monitoring and switching system that employs state-of-the-art Quality Estimation methods.  The overall system produces highly fluent translations in most cases, and reverts to a less fluent but hallucination-free mode when needed to virtually eliminate incorrect outputs.  The Phase II effort will include an end-to-end demonstration of the process, resulting in a deployed MT system for a group of related low-resource language.  The complete deployed system will include a general-purpose translation capability for the selected languages as well as one or more variants optimized for specific Army needs, e.g., optimized for a specific domain of interest to the Army, and/or optimized for use in a specific downstream task such as stance detection.