Adapting Pre-Trained Models and Leveraging Targeted Multilinguality for Under-Resourced and Endangered Language Processing

Abstract

Advances in Natural Language Processing (NLP) over the past decade have largely been driven by the scale of data and computation used to train large neural network-based models. However, these techniques are inapplicable to the vast majority of the world's languages, which lack the vast digitized text datasets available for English and a few other very high-resource languages. In this dissertation, we present three case studies for extending NLP applications to under-resourced languages. These case studies include conducting unsupervised morphological segmentation for extremely low-resource languages via multilingual training and transfer, optimizing the vocabulary of a pre-trained cross-lingual model for specific target language(s), and specializing a pre-trained model for a low-resource language family (Uralic). Based on these case studies, we argue for three broad, guiding principles in extending NLP applications to under-resourced languages. First: where possible, robustly pre-trained models and representations should be leveraged. Second: components of pre-trained models that are not optimized for new languages should be substituted or substantially adapted. Third: targeted multilingual training provides a middle ground between the lack of adequate data to train models for individual under-resourced languages on one hand, and the diminishing returns of ``massively multilingual'' training on the other.