Syntactic Transfer for Low-Resource Machine Translation with Contextual Parameter Generation

Abstract

The advent of large pretrained models has led to paradigm-shifting improvements throughout natural language processing. For many tasks, state-of-the-art results are now achieved by taking one of these large pretrained models and adapting it in some way for use on the desired task. While this approach has been successful on a broad range of tasks, that success is not evenly distributed within tasks—most of the gains are in high-resource settings, i.e., tasks and languages for which there is a large amount of labeled data available. Some tasks—and many languages—lack sufficient labeled data for these approaches to work well. Recently, there has been much interest in methods that could potentially close this gap and improve performance in low-resource settings. In this work, I demonstrate a novel method for adapting large pretrained models that involves dynamically generating additional parameters for the model based on an informative representation of the task and show that this method works especially well on the task of low-resource machine translation.