This paper proposes a zero-shot grammar competency estimation framework leveraging unlabeled data and Large Language Models (LLMs) to generate pseudo-labels for training a transformer-based model. It addresses the challenges of limited labeled data in grammar assessment, particularly for spoken language.
