ENHANCING TRANSLATION SYSTEMS FOR LOW-RESOURCED SETTINGS

There are around 7000 languages that are alive worldwide; among them, only 50-200 languages are well-resourced. In many regions of the world, there are languages and dialects with limited resources, which are at risk of disappearing due to sociopolitical problems and a lack of attention. As a result, those communities need more technologies that could significantly impact areas such as education, healthcare, and emergency response. Low-resource translation addresses this problem by developing translation systems and tools for languages and dialects with limited resources to revive those languages. The traditional translation method requires the collection of bilingual text-to-text resources. Low-resourced languages lack these resources because more than 3000 languages are oral; even if a writing system exists, most people are illiterate; even if literate, most communities do not have a robust online presence, which is the primary source of data acquisition; even if they have an excellent online presence, the resource is often too noisy. This work aims to solve specific problems that arise in the low-resource scenario due to the minimal annotated resources available for that particular scenario. The problems that we tackle cover different sectors of translation:• Translating L2-speaker variation: The translation model should understand the L2-speaker’s errors (e.g., grammatical and spelling). • Translating dialectal variation: The translation model should be able to handle dialectal input (or output) to/from other languages. • Translating low-resourced African languages: Creating translation models for low-resourced languages focusing on African languages (e.g., Zulu, Wolof, Hausa, Igbo, Bemba). • Translating terminology: The translation model should generate domain-specific terminology (e.g., Medical, Law). • Translating low-resourced languages using a dictionary: Creating translation models using other clean, available data sources (e.g., dictionary, books) • Translating speech-to-text: Half of the low-resourced languages in the world are oral, and most of the speech data are noisy. Filtering and augmenting speech data are crucial for creating translation systems.