Using Meta-Learning to address the task-agnostic problem in Natural Language Understanding
Human beings can generalize from a single example of a task to similar tasks. However, computers need to effectively see thousands of examples to learn to generalize to similar tasks. To learn computers need to be given labeled datasets, which include both correct and incorrect examples, properly labeled. The limited amount of available labeled datasets makes it difficult for machine learning to achieve human-level performance. Meta-Learning, which is known as "learning to learn", can address this difficulty. This means that Meta-Learning enables learning with a few training samples and enables adaptation across domains or in a single constantly changing domain. Meta-Learning has been explored and used in robotics and computer vision, but it has not been sufficiently explored in natural language processing. This research aims to explore how to use Meta-Learning in addressing task-agnostic problems in Natural Language Understanding.